Deep Learning For Amharic Text-ImageRecognition

Deep Learning for Amharic Text-Image
Recognition: Algorithm, Dataset and Application
Thesis
approved by
the Department of Computer Science
Technische Universität Kaiserslautern
for the award of the Doctoral Degree
Doctor of Engineering (Dr.-Ing.)
to
Birhanu Hailu Belay
Date of Defense: 10.02.2021

Dean: Prof. Dr. Jens Schmitt
Chairperson of the PhD Committee:
Prof. Dr. Ralf Hinze
Reviewer: Prof. Dr. Didier Stricker
Reviewer: Prof. Dr. Marcus Liwicki
D 386
Abstract
Optical Character Recognition (OCR) is one of the central problems in pattern recognition. Its
applications have played a great role in the digitization of document images collected from het-
erogeneous sources. Many of the well-known scripts have OCR systems with sufficiently high
performance that enables OCR applications in industrial/commercial settings. However, OCR sys-
tems yield very-good results only on a narrow domain and very-specific use cases. Thus, it is still
a challenging task, and there are other exotic languages with indigenous scripts, such as Amharic,
for which no well-developed OCR systems exist.
As many as 100 million people speak Amharic, and it is an official working language of Ethiopia.
Amharic script contains about 317 different alphabets derived from 34 consonants with small changes.
The change involves shortening or elongating one of its main legs or adding small diacritics to the
right, left, top, or bottom of the consonant character. Such modifications lead the characters to have
similar shapes and make the recognition task complex, but this is particularly interesting for charac-
ter recognition research. So far, Amharic script recognition models are developed based on classical
machine learning techniques, and they are very limited in addressing the issues for Amharic OCR.
The motivation of this thesis is, therefore, to explore and tailor contemporary deep learning tech-
niques for the OCR of Amharic.
This thesis addresses the challenges in Amharic OCR through two main contributions. The first
contribution is an algorithmic contribution in which we investigate deep learning approaches that
suit the demand for Amharic OCR. The second is a technical contribution that comprises several
works towards the OCR model development; thus, it introduces a new Amharic database consisting
of collections of images annotated at a character and text-line level. It also presents a novel CNN-
based framework designed by leveraging the grapheme of characters in Fidel-Gebeta (where Fidel-
Gebeta consists of the full set of Amharic characters in matrix structure) and achieves 94.97%
overall character recognition accuracy.
In addition to character level methods, text-line level methods are also investigated and devel-
oped based on sequence-to-sequence learning. These models avoid several of the pre-processing
stages used in prior works by eliminating the need to segment individual characters. In this design,
we use a stack of CNNs, before the Bi-LSTM layers and train from end-to-end. This model out-
performs the LSTM-CTC based network, on average, by a CER of 3.75% with the ADOCR test
set. Motivated by the success of attention, in addressing the problems’ of long sequences in Neural
Machine Translation (NMT), we proposed a novel attention-based methodology by blending the
attention mechanism into CTC objective function. This model performs far better than the existing
techniques with a CER of 1.04% and 0.93% on printed and synthetic text-line images respectively.
Finally, this thesis provides details on various tasks that have been performed for the development
of Amharic OCR. As per our empirical analysis, the majority of the errors are due to poor annotation
of the dataset. As future work, the methods proposed in this thesis should be further investigated
and extended to deal with handwritten and historical Amharic documents.
Acknowledgments
First, I would like to thank Prof. Dr. Didier Stricker for allowing me to conduct my Ph.D. research
at the Augmented Vision lab, DFKI-Kaiserslautern. Without his ongoing support, supervision,
feedback, and the freedom to investigate different approaches, this thesis would not have been
possible. Further, I would like to thank Prof. Dr. Marcus Liwicki for becoming my cosupervisor, his
deep insights, invaluable feedback, fruitful discussions, and many great ideas are the cornerstones
of this research work. I would like to thank Dr. Million Meshesha and Dr. Gebeyehu Belay, my
mentors in the Ph.D. program for their guidance and feedback on my research.
I would like to thank Tewodros, who led to the first conceptual ideas for this thesis. Not only the
long technical discussions but also your continuous motivation and support have always been my
asset. Thank you for taking the time for the many fruitful discussions that not only challenged me
but also helped me to develop many of the ideas behind this thesis and teaching me the importance
of hard work. Then, Fatemeh, you and Teddy were always there to facilitate and made things
easy during my conference registration wherever and whenever it is. Special thanks to Reza, who
initially helped me a lot in understanding various ideas regarding my topic and letting me know
and use the OCRopus tool. Many thanks also to Kina for her valuable feedback and proofreading
towards the end of this thesis. I would like to acknowledge David, who is always there to maintain
and provide a solution immediately when my computing infrastructure faces any problem. I would
like to acknowledge Dr. Jason for his support during the thesis write-up. Whenever I drop an email
to him, he always replied to the point and immediately. Last but not least, special thanks to Leivy,
who helped me a lot from the very beginning of my arrival at DFKI until I was officially accepted
as a Ph.D. student at TU-Kaiserslautern. Further, I would like to thank Keonna, Nicole, and Eric
for helping me with various administrative issues.
I would like to thank Abrham Gebru, for spending countless hours together. Especially, every
random weekend trip, we had to Frankfurt to buy Injera. Even though it was my first winter semester
in Germany, he made me feel like it’s summer. My best wishes to him. I would like to thank Dr.
Tsegaye, Dr. Dereje, and Dr. Mulugeta, for all the times we spent in KL. I would especially like to
thank Dere, for being there for me during the pandemic, he is a very sociable person and we had a
good friendship and spent lots of time together and he was a very great chef of our kitchen.
It is a pleasure to express my gratitude wholeheartedly to my wife, Masti, for all her love, pa-
tience, and having my back in stressful times. She has also waited and stayed with me till midnight,
and she is always there to encourage me when I work the whole night. My daughter, Nissan, and
my son, Zephaniah, are not only my kids, but you are also my friends, who understand me when I
work and bring out the best in me when I got bored. I am also very thankful to my brother, Workye
Hailu, for caring in all my ways. My mother, Enatalem, for encouraging and supporting me in all
my studies starting from early schools. I am also very grateful to my father, Hailu Belay, and the
rest of all my families and friends who have a significant contribution and help me in one way or
another during my research work.
Birhanu Hailu
Contents
Abstract ix
List of figures xii
List of tables xiii
Abbreviations xvi
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Statement of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 List of Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Trends in OCR Research and Development 13

2.1 Classical OCR Approach . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Modern OCR Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Deep Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 Convolutional Neural Networks . . . . . . . . . . . . . . . . . 18
2.3.2 Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . 19
2.4 Sequence-to-sequence Learning for OCR . . . . . . . . . . . . . . . . 20
2.4.1 CTC-based Sequence Learning . . . . . . . . . . . . . . . . . . 20
2.4.2 Attention-based Sequence Learning . . . . . . . . . . . . . . . 22
2.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Amharic Script and Writing System 29

3.1 Amharic Script and its Genetic Structure . . . . . . . . . . . . . . . . . 29
3.2 Orthographic Identity of Amharic Characters . . . . . . . . . . . . . . 31
3.2.1 Features of Amharic Characters . . . . . . . . . . . . . . . . . 33
3.3 Challenges for the OCR of Amharic Script . . . . . . . . . . . . . . . . 35
vii
viii Contents
3.4 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4 Amharic OCR and Baseline Datasets 41

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Dataset for OCR Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Database for Amharic OCR . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.1 Amharic Character Image Dataset . . . . . . . . . . . . . . . . 44
4.3.2 Amharic Text-line Image Dataset . . . . . . . . . . . . . . . . 45
4.4 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5 Recognition of Amharic Characters 53

5.1 Amharic Character Recognition Using CNNs . . . . . . . . . . . . . . 53
5.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.1.3 Research Methods . . . . . . . . . . . . . . . . . . . . . . . . 56
5.1.4 The Proposed CNN Architecture . . . . . . . . . . . . . . . . . 57
5.1.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . 58
5.1.6 Recognition Performance Comparison . . . . . . . . . . . . . . 60
5.1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2 Amharic Character Grapheme-based OCR . . . . . . . . . . . . . . . . 61
5.2.1 Character-level OCR . . . . . . . . . . . . . . . . . . . . . . . 62
5.2.2 Research Methods . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.3 Multi-Task Learning . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.4 The Proposed FCNN Architecture . . . . . . . . . . . . . . . . 65
5.2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.3 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6 Amharic Script Recognition using CTC Sequence Modeling 71

6.1 Amharic Text-image Recognition with LSTM-CTC Networks . . . . . 71
6.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.1.2 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.1.3 The Proposed Model Architecture . . . . . . . . . . . . . . . . 73
6.1.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.1.5 Analysis of the Result . . . . . . . . . . . . . . . . . . . . . . 78
6.1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 Deep CNN-RNN Hybrid Networks for Amharic OCR . . . . . . . . . . 79
6.2.1 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2.2 The Proposed Hybrid CNN/LSTM Network Architecture and
Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2.4 Discussion and Analysis of Results . . . . . . . . . . . . . . . 85
Contents ix
6.2.5 Recognition Performance Comparison . . . . . . . . . . . . . . 86

6.2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.3 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7 Attention-based Sequence Learning for Amharic OCR 89

7.1 Attention-based Encoder-Decoder Model for Amharic OCR . . . . . . 89
7.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7.1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.1.3 The Proposed Attention-based Network . . . . . . . . . . . . . 92
7.1.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
7.2 Blended Attention-CTC for Amharic Text-image Recognition . . . . . 97
7.2.1 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2.2 The Proposed Blended Attention-CTC Model . . . . . . . . . . 99
7.2.3 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.2.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.3 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
8 Conclusions and Recommendations for Future Work 105

8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
8.2 Recommendations for Future Work . . . . . . . . . . . . . . . . . . . . 108
Appendices 111
A.1 Attention Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
A.1.1 Generic Model of Attention Mechanism . . . . . . . . . . . . . 114
A.1.2 Types of Attention and Seq2seq Model . . . . . . . . . . . . . 115
A.1.3 Basic Parameters of Attention-based Networks . . . . . . . . . 122
Bibliography 124
List of Figures
1.1 A generic Document Image Analysis (DIA) flow diagram . . . . . . . 2

1.2 Sample Amharic and Ge’ez document collections . . . . . . . . . . . . 4
2.1 Character segmentation process and segmentation problems on sample

Amharic text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Illustration of CTC-based sequence learning . . . . . . . . . . . . . . . 22
2.3 Illustration of Attention-based sequence learning . . . . . . . . . . . . 25
2.4 A typical attention model . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1 Demographic distribution of Amharic speakers . . . . . . . . . . . . . 30

3.2 Family of Amharic language . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 List of basic Amharic characters . . . . . . . . . . . . . . . . . . . . . 33
3.4 Other symbols of Amharic Fidel . . . . . . . . . . . . . . . . . . . . . 34
3.5 Sample Amharic scripts . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1 Sample synthetically generated Amharic character images . . . . . . . 44

4.2 Fidel Gebeta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Synthetic Amharic text-line image generation pipelines . . . . . . . . . 46
4.4 A sample scanned page of Amharic document . . . . . . . . . . . . . . 47
4.5 A typical diagram of OCRopus page segmentation-engine . . . . . . . 48
4.6 Amharic characters and punctuation marks . . . . . . . . . . . . . . . . 49
4.7 Text-line image length normalization . . . . . . . . . . . . . . . . . . . 49
4.8 Sample Amharic text-line images from ADOCR database . . . . . . . . 50
5.1 A CNN-based OCR model for Amharic character image recognition . . 58

5.2 Training versus validation loss of CNN-based OCR model . . . . . . . 59
5.3 Shape changes in row-column structure of Fidel-Gebeta . . . . . . . . . 64
5.4 Factored CNN architecture: . . . . . . . . . . . . . . . . . . . . . . . . 65
5.5 A typical diagram that shows the prediction of FCNN model . . . . . . 67
6.1 The proposed network architecture . . . . . . . . . . . . . . . . . . . . 75

6.2 Sample text-line image for illustration of Amharic script . . . . . . . . 76
xi
xii List of Figures
6.3 Training versus validation loss of an LSTM-CTC model . . . . . . . . 77

6.4 Sample miss-recognized characters in each text-line image of the test
dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.5 Sample Amharic text-line images from ADOCR database . . . . . . . . 81
6.6 The proposed CNN-LSTM model . . . . . . . . . . . . . . . . . . . . 82
6.7 Sample wrongly annotated images and GT from the test dataset . . . . . 86
6.8 Learning loss comparison . . . . . . . . . . . . . . . . . . . . . . . . . 87
7.1 Attention-based Encoder-Decoder model for Amharic OCR . . . . . . . 94

7.2 Sample predicted text images using attention-based OCR model . . . . 96
7.3 The proposed blended Attention-CTC model . . . . . . . . . . . . . . . 100
7.4 The training & validation losses of model trained with different network
settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.5 Sample text-line image with the corresponding predicted and ground-
truth texts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
A.1 General setting of Attention model . . . . . . . . . . . . . . . . . . . . 114

A.2 Seq2seq model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
A.3 A typical architecture of Soft attention . . . . . . . . . . . . . . . . . . 118
A.4 Attention model architecture based on the work of Luong . . . . . . . . 121
List of Tables
5.1 Character level recognition accuracy (out of 100%) and performance

comparison with related research works. . . . . . . . . . . . . . . . . 60
5.2 Performance of prior works (CNN) and proposed method(FCNN) . . . 67
6.1 The detail of Amharic text-line images database . . . . . . . . . . . . . 74

6.2 A summary of experimental results (CER) of LSTM-CTC model . . . . 77
6.3 Convolutional network layers of the proposed model and their corre-
sponding parameter values for an input image size 32 × 128 × 1. . . . 83
6.4 The recurrent network layers of the proposed model with their corre-
sponding parameter values . . . . . . . . . . . . . . . . . . . . . . . . 84
6.5 Experimental results of CNN-LSTM OCR model in CER (%). . . . . . 85
6.6 Comparison of test results (CER). . . . . . . . . . . . . . . . . . . . . 87
xiii
Abbreviations
OCR Optical Character Recognition

DIA Document Image Analysis
NLP Natural Language Processing
MTL Multi-Task Learning
NMT Neural Machine Translation
EOTC Ethiopian Orthodox Tewahedo Church
ENALA Ethiopian National Archives and Library Agency
PR Pattern Recognition
RNN Recurrent Neural Network
LSTM Long Short-Term Memory
CNN Convolutional Neural Network
Bi-LSTM Bidirectional LSTM
CTC Connectionist Temporal Classification
HMM Hidden Markov Models
ADOCR Amharic Database for OCR
ANN Artificial Neural Network
SVM Support Vector Machine
FCNN Factored CNN
CRNN Convolutional RNN
CER Character Error Rate
FCN Fully Connected Network
BACN Blended Attention CTC
AED Attention based Encoder-Decoder
xv
Chapter 1
Introduction
The history of Optical Character Recognition (OCR) is traced back to the

1809s [Her82, DT+ 14] when the first reading machine for the visually im-
paired was patented. Since then, it is being applied throughout the spectrum
of industries, resulting in the revolutionization of the document management
process. The application of character recognition is not limited to converting
image data to text but it has also proliferated with various businesses including
banking, legal industries, health care systems, institutional repositories and
digital libraries, all over the world. In previous years, there were a number of
problems that affected the implementation of OCR due to its’ computationally
intensiveness and training data shortage. Further, early versions of OCR had
to be trained with images of each character and was limited to recognizing a
single font at a time. However, nowadays, the introduction of faster Graph-
ical Processor Units (GPU) and implementation of deep learning algorithms
have enabled lots of exciting new features such as automatic multiple feature
extraction, scalability for complex images and documents, which pushed the
achievements of sequence-to-sequence tasks and OCR applications further.
Optical Character Recognition (OCR) refers to a set of computer vision

problems that are employed to convert images of machine-printed texts or
hand-written text images to machine-readable text in a format that a com-
puter can process, store, and edit as a text file; or as a part of data entry for
further manipulation [BT14]. Digitization of textual resources and hard-copy
documents including books, newspapers, magazines, cultural and religious
archives have been underway for decades to promote world-wide accessibil-
ity of these resources. Nowadays, documents are processed with comput-
ers, however, it is unclear whether the computer has decreased or increased
1
2 Introduction
Layout-analysis
Preprocessing Column detection
Image / non-image
Noise removal detection Segmentation
Skew detection Title, keyword and
Binarization line detection Line, word and/or
character segmentation
Feature-extraction
Hand-crafted
Automatic feature
Document Image
Models
Recognition
Post processing
Dictionary correction
Model Language model
Alignment correction
corrected text
Document Texts
Figure 1.1: A generic Document Image Analysis (DIA) flow diagram . This
generic DIA flow diagram can be broken down into four major
steps: The first step is image preprocessing which helps to en-
hance the quality of the image and it involves broad range of
imaging functions such as image rotation, binarization and de-
skewing. The second document analysis step defines the areas for
text recognition and delivers information about layout and for-
matting elements of each individual page as well as structure of
the document as a whole. The actual texts are predicted at recog-
nition step. Then the OCR errors are corrected and the model is
updated at the post-processing stage.
the amount of paper-based documents. A large number of documents, hav-

ing different contents and written with different scripts, are still processed,
archived, and disseminated in paper format [KJEY20]. During recognition,
once a printed and/or handwritten document has been captured optically us-
ing digital camera, scanner or any other optical means, then the digital image
goes to the following major steps of Document Image Analysis (DIA) for fur-
ther processing and recognition. First, it goes to the preprocessing stage and
then it goes to the feature extraction stage. Finally, the features obtained at
the feature extraction step are utilized for the development of the recognition
model. Figure 1.1 depicts the generic process of DIA and basic activities in
each step.
3
• Preprocessing stage: In this stage the quality of the images are en-
hanced employing different image preprocessing techniques and the
data of interest are located.
• Feature extraction: The distinctive characteristics of the image, useful

for recognition, are captured in the feature extraction stage. Depend-
ing on the type of algorithm employed for feature extraction, document
layout correction and image segmentation could be applied at different
levels.
• Classification stage: Here, the extracted feature vectors from the pre-
vious stage are processed to recognize character, word, or text-lines.
• Post-processing: At this step, the recognition errors are corrected based

on contextual information from the language-model or dictionary. As a
result, post-processing helps to update and/or improve the recognition
performance of the OCR model.
The applications of OCR have been widely used and implemented for the
digitization of various documents, written in both Latin and Non-Latin scripts,
ranging from historical to modern documents [Mes08, BHL+ 19a, CPT+ 19,
IFD+ 19]. Researchers achieved a better recognition accuracy and even most
of these scripts, now, have workable and commercial off-the-shelf OCR ap-
plication including the functionality of the ground truth generator from the
exiting printed text so as to train the second model [Bre08]. However, OCR
gives a better recognition accuracy only on a narrow domain and very specific
use cases. Besides, other multiple indigenous scripts are underrepresented in
the area of Natural Language Processing (NLP) and Document Image Analy-
sis (DIA). Hence, researchers are motivated towards the development of either
a generic multi-script OCR model or an OCR model for specific scripts.
Dated back to the 12th century, most of the historical and literary docu-
ments in Ethiopia are written and documented using Amharic script [Mey06].
In parallel, there are some documents written in Ge’ez script which share the
same writing system and characters as Amharic script. Since then, multiple
documents containing various contents are stored in different places such as
Ethiopian Orthodox Tewahdo Churches, museums, monasteries, public and
academic libraries in the form of correspondence letters, magazines, newspa-
pers, pamphlets, and books [Hag15]. These documents contain information
1
S. the on-line database Mazgaba seelat (Ruba Kusa); on historical carpets of Rubakusa,
s. Gervers 2004:292-93: Grant ES. Ethio-SPaRe Cultural Heritage of Christian Ethiopia:
Salvation, Preservation and Research.
4 Introduction
Figure 1.2: Sample Amharic and Ge’ez document collections. The top im-
age of this figure is Ge’ez document 1 while the others are Amharic
documents. Each document image depicts variations of fonts,
styles, and layouts of Amharic script that are limiting the im-
plementation of a robust OCR system. The characters of Ge’ez
are a subset of Amharic characters; thus, from the OCR perspec-
tive, a recognition model of Amharic script can be directly applied
and/or extended to Ge’ez script recognition.
related to religion, history, literature, politics, economics, philosophy, aca-

demic, traditions, nature, and cultural pieces of evidence of the people of
Ethiopia. With a digitization campaign, many of these manuscripts are col-
lected from various sources. However, they are still preserved in a manual
catalog and/or scanned copies of them in Microfilm format [Wio06]. Sample
Amharic and, its ancestor, Ge’ez document collections are shown in Figure
1.2. These documents consist of representative images of historical, modern,
handwritten and machine-printed documents with varieties of writing style
and document layouts.
Nowadays, document processing and preservation has been given much

attention by many researchers from the field of linguistics and social science
[MM18, AMT+ 18]. In this regard, the Ethiopian Orthodox Tewahedo Church
(EOTC) and Ethiopian National Archives and Library Agency (ENALA) has
1. Motivation 5
played a crucial role in shaping national strategy and supporting digital in-
novation by the public organization with their implementation of digital ini-
tiatives [CBC19, NAL]. To support librarian, historian, and other experts in
the humanity studies, academic and industrial researchers have been attempt-
ing to develop OCR model to digitize textual-resources and documents for
various scripts. However, the nature of the script, such as type of fonts, repre-
sentation of encoding, quality of printing and support from operating systems
affect document standardization, causing the languages to have diverse fea-
tures in nature [Mes08]. Consequently, these issues add to the complexity of
the design and implementation of document image recognition systems to a
great extent for many scripts including Amharic script.
1.1 Motivation
The main motivation for this thesis is the research gap in the OCR develop-
ment for Amharic script. Amharic follows a unique writing system that has its
specific linguistic composition and indigenous characters. Amharic language
has a rich collection of documents that are collected from various sources over
centuries that should be digitized for further research and knowledge dissem-
ination. Consequently, there is a need to develop an OCR model for Amharic
script. However, despite the need for developing the OCR model, there is
a very limited research attempt for developing a robust and automatic OCR
system for Amharic script.
So far, the research attempts made for Amharic OCR development are not
well organized. These attempts are also implemented based on classical ma-
chine learning algorithms which require lots of effort for prepossessing and
feature extraction. The type and number of characters considered in the train-
ing set are not enough to represent the occurrence of characters even in com-
mon Amharic documents. In addition, the performance of the OCR model in
general and specifically the OCR models developed for Amharic script recog-
nition are affected by the nature and varieties of document images. The un-
availability of organized datasets for training is also one of the limiting factors
in Amharic OCR development. Hence, a large annotated real-world dataset is
required to train OCR systems and improve their performance. However, it is
a very challenging and time-consuming task to manually annotate real-world
documents. Therefore, a method is needed, which can automatically generate
ground truth for character or text-line images that can represent real-world
artifacts of Amharic documents. This motivated the author to introduce a par-
6 Introduction
ticular and simple method for the preparation of printed texts and automatic
ground-truth generation of synthetic Amharic text-line images.
1.2 Statement of the Problem
The present growth of document digitization and changes in the world of the
linguistic landscape demands an immediate solution for enabling informa-
tion access for everyone. This requires research in the area of natural lan-
guage processing and document image analysis. In parallel, the demands of
language technology are increasing for resource-limited languages, to break
language barriers. Amharic is one of these resource-limited languages with
about 100 million speakers around the world. It has an indigenous script
with its own alphabets. Amharic is rich in historical documents collected
over centuries. However, from an OCR perspective, there is only limited re-
search effort for Amharic scripts. The scripting nature of the language and
lack of data availability poses many challenges in document image analysis
and recognition tasks.
For decades, OCR was the only way to turn printed documents into data
that could be processed by computers, and it remains the tool of choice for
converting paper documents into editable data that can be processed further.
Ultimately, OCR will continue to be a valuable tool for filling in gaps of hu-
manity and the digital world. Since the truly paperless business doesn’t yet
exist, data extraction is still a useful tool and that can augment the develop-
ment of OCR technology to minimize the scope for errors.
Deep learning-based approaches have improved over the last few years, re-
viving an interest in the OCR problem, where neural networks can be used
to combine the tasks of localizing texts in an image along with understand-
ing what the texts are. Using CNN architectures, Attention mechanisms and
RNNs have gone a long way in this regard. However, to the best of our knowl-
edge, researches are done on Amharic OCR so far is in an uncoordinated
manner and none of the researchers have taken the advantage of deep learn-
ing techniques such as sequence-to-sequence learning instead segmentation-
based traditional machine learning techniques have been employed nearly in
all of these attempts. In addition, in the literature, attempts to Amharic OCR
neither shown results on a large dataset nor considering all possible charac-
ters used in Amharic script. Therefore, this work aims to explore the problem
of OCR in Amharic script and solve some of these problems by applying
3. Research Objectives 7
state-of-the-art deep learning algorithms. In parallel, we develop a baseline

Amharic database so as to address the issue of the dataset.
1.3 Research Objectives
Given the increasing need to digitize documents and its impact on accessibil-
ity of indigenous knowledge, this thesis aims to construct a robust OCR model
for Amharic script recognition using contemporary deep learning algorithms.
This thesis also aims to reduce the research gap in Amharic language and
to develop a baseline database for Amharic OCR which can help researchers
for further improvement of the OCR system. To achieve the objective of this
thesis, we pose the following theoretical hypotheses and we are going to test
them through an empirical investigation.
1. The performance of OCR model for basic Amharic character recogni-

tion can be improved if we develop a model of multi-task networks by
leveraging the grapheme of characters on Fidel-Gebeta.
2. The ability of RNNs to process contextual information enables them

to learn sequence-to-sequence mapping; thus, stacking CNNs before
RNNs as a feature extractor and employing CTC alignment making
them an appropriate choice for the task of text-image recognition.
3. In NMT, Attention enhances the ability of the networks in extracting the

most relevant feature for each part of the output sequence; thus, it suits
and performs better if we extend Attention mechanism for Amharic
text-line image recognition.
4. Given that the second and third hypotheses are correct, then using the
best of both worlds and blending the capability of Attention mechanism
into CTC objective function boosts the recognition performance of the
OCR model.
1.4 Contributions
In this thesis, we focus on designing an OCR model for Amharic script. Some
of the main conceptual and technical contributions of this thesis are summa-
rized as follows:
8 Introduction
• A well structured and standard database of basic Amharic character im-

ages is developed in this research work. The constructed database con-
sists of 80,000 character images and it has the potential to be used as a
benchmarking resource for Amharic character recognition.
• Created the first publicly available Amharic text-line image dataset.

The dataset is a collection of 337,337 synthetically generated and printed
text-images. It is also evaluated by state-of-the-art OCR techniques
such as LSTM-CTC, and Attention-based encoder-decoder networks.
• Propose a novel Approach for basic Amharic character recognition. To

the best of our knowledge, it is the first work designed as a multi-task
architecture by leveraging the structure of basic Amharic characters
grapheme’s on Fidel Gebeta.
• Designed an OCR for recognizing Amharic scripts at text-line level.

These OCR models are extensively tested on both synthetically gener-
ated and printed images and promising results are obtained.
• Adoption of synthetic database generation methodology and propose

techniques for printed text-image dataset preparation for Amharic script.
• The basic idea of Attention mechanism, on neural machine translation,

is extended to the OCR of Amharic script where an Attention based
encoder-decoder networks recognize text-line images. With this design
we achieve a recognition efficiency comparable with CTC-based recur-
rent networks.
• Propose a novel sequence-to-sequence learning method by blending the

concept of Attention mechanism directly into CTC. This technique is
evaluated with the test sets of ADOCR database and outperforms other
state-of-the-art OCR techniques by a large margin.
• A detail analysis of the results of experiments carried out with different

deep learning techniques is presented. Compared to the classical ma-
chine learning based OCR approaches, the deep learning based OCR
approaches yield the best reported results for Amharic scripts recogni-
tion and has fewer preprocessing steps.
5. Organization of the Thesis 9
1.5 Organization of the Thesis
This thesis is organized into eight chapters. The first chapter provides an
overview of the general background of OCR in general and particularly, for
Amharic script recognition. The motivation behind the present work, state-
ment of the problem, research objectives, list of publications from this thesis,
and the major contributions are also briefly described. In chapter two, we
present the trends in OCR research and various methodologies that have been
emerged over decades of OCR research. Both traditional and state-of-the-art
OCR approaches that were and have been applied in DIA and OCR research
are also presented. Specifically, it is noted on sequence-to-sequence learning
techniques such as Attention mechanism and CTC-based recurrent networks
have been utilized, in this thesis, for Amharic scripts.
Chapter three introduces Amharic languages and presents background in-

formation about Amharic, the demographic distribution of its’ speakers, the
genetic structure of Amharic script, character sets, and writing system. Fur-
ther, related issues that are hindering the development of a reliable Amharic
OCR system and unique features of Amharic script have been described. This
chapter also gives a summary of related works and highlights that the unavail-
ability of training datasets is the major obstacle to Amharic OCR research.
Chapter four is about the dataset used in this thesis; then it introduces and
describes the details of an Amharic database called ADOCR which contains
both printed and synthetically generated images. The dataset is organized into
two levels; at the character level and text-line level image. Data collection and
synthetic data generation techniques are also described.
Chapter five discusses CNN-based Amharic character recognition models

and experimental results. Unlike the usual Amharic character recognition
model, characters are recognized based on their grapheme on Fidel-Gebeta
employing a multi-task convolutional neural network model and achieved bet-
ter character recognition results. In this chapter, multi-task learning paradigms,
and related research works done on Amharic character recognition are also
presented in detail.
Chapters six and seven provide the systems architecture of various sequence-
to-sequence learning models. Chapter 6 focuses on using LSTM as a sequence-
to-sequence mapping together with CTC. CNN’s were also covered and em-
ployed as a feature extractor module by stacking before recurrent networks.
Further, various recurrent network-based approaches are incorporated in the
10 Introduction
learning processes and experimental results are presented to show their appli-
cability and performance in recognition of Amharic text-images.
Attention mechanism has been given more attention in NLP tasks for neu-
ral machine translation. Based on the idea of neural machine translation, an
Attention-based encoder-decoder network is proposed in Chapter 7. In the
first part of this Chapter, an end-to-end Amharic OCR model is proposed us-
ing LSTM networks, without CTC alignment, as encoder and decoder unit.
Attention is embedded between the encoder and decoder LSTMs to allow the
decoder to attend to different parts of the source sequence and to let the model
learn how to generate a context vector for each output time step. This model
gives state-of-the-art results on ADOCR test datasets. An extension of this
work, a blended Attention-CTC network formulated by directly taking the ad-
vantage of Attention mechanism and CTC network, is reported in the second
part of this chapter. The performance of this blended attention-CTC network
is evaluated against the ADOCR test sets and it performs significantly better
than the other sequence-to-sequence models.
Finally, a summary of this thesis, concluding remarks, and guidelines for

future work are provided in chapter eight.
1.6 List of Publications
Part of the work presented in this thesis has been accepted and/or presented
in peer-reviewed conferences or journals. The list of the publications derived
from this thesis work are provide as follows.
Journals
1. Birhanu Hailu Belay, Tewodros Habtegebrial, Million Meshesha, Mar-

cus Liwicki, Gebeyehu Belay, Didier Stricker, ”Amharic OCR: An End-
to-End Learning”. MDPI-Applied Sciences. 2020; 10(3):1117.
Conferences and Preprints
1. Birhanu Hailu Belay, Tewodros Habtegebrial, Marcus Liwicki, Gebeyehu

Belay, Didier Stricker, ”Factored Convolutional Neural Network for
6. List of Publications 11
Amharic Character Image Recognition”, 26th IEEE International Con-

ference on Image Processing (ICIP), Taipei, Taiwan, 22–25 September
2019; pp. 2906–2910.
Belay, Didier Stricker, ”Amharic Text Image Recognition: Dataset, Al-
gorithm and Analysis”, 15th International Conference Document Anal-
ysis and Recognition (ICDAR), Sydney, Australia, 20–25 September
2019; pp. 1268–1273.
3. Birhanu Hailu Belay, Tewodros Habtegebrial, Gebeyehu Belay, Didier
Stricker, ”Multi-font Printed Amharic Character Image Recognition:
Deep Learning Techniques”, 6th EAI International Conference on Ad-
vances of Science and Technology: ICAST 2018, Bahir Dar, Ethiopia,
October 5-7, 2018, Proceedings, vol. 274, p. 322. Springer, 2019
4. Birhanu Hailu Belay, Tewodros Habtegebrial, Didier Stricker, ”Amharic
character image recognition”, 18th IEEE International Conference on
Communication Technology (ICCT), Chongqing, China, 8–11 October
2018; pp. 1179–1182
Belay, Didier Stricker, ”Blended Attention-CTC Network Architecture
for Amharic Text-image Recognition, 10th International Conference on
Pattern Recognition Applications and Methods (ICPRAM 2021)
6. Birhanu Hailu Belay, Tewodros Habtegebrial, Gebeyehu Belay, Mil-
lion Meshesha, Marcus Liwicki, Didier Stricker, ”Learning by Injec-
tion: Attention Embedded Recurrent Neural Network for Amharic Text-
image Recognition”. MDPI-Preprints 2020
Chapter 2
Trends in OCR Research and
Development
This chapter provides a highlight of the existing methods and trends in OCR
technology. Technically, OCR comprises numerous procedures and tech-
niques which have been emerged over a century of research. The rapidly
growing and availability of computing power has facilitated the use of so-
phisticated and diversified approaches for OCR development. The following
sections of this chapter have acknowledged multiple approaches of OCR re-
search which can be broadly categorized into two groups. The first category,
the classical OCR approach, is commonly called segmentation-based OCR
which involves the segmentation of a page into paragraphs, then into text-
lines, followed by words, and finally into individual characters. A character
recognizer would recognize the segmented characters one by one. Section 2.1
describes various methods developed for this type of OCR approach. The sec-
ond OCR approach, which is called segmentation-free OCR in literature, and
the modern OCR approach in this thesis. Section 2.2 describes this OCR ap-
proach in detail. Section 2.3 describes the overview of deep neural networks
and section 2.4 presents state-of-the-art and emerging trends of segmentation-
free OCR approach and Seq2Seq tasks in general. Since recognition in seg-
mentation free approach is done at the word or at text-line level, and it avoids
the task of extracting characters from text-lines and minimizes the efforts re-
quired for preprocessing and feature extraction.
13
14 Trends in OCR Research and Development
2.1 Classical OCR Approach
The classical OCR approach is commonly called the segmentation-based ap-

proach. Segmentation is the process of partitioning a digital image into multi-
ple manageable segments in general and specifically for OCR, it is the process
of segmenting document images into characters. The goal of segmentation is
to simplify and/or change the representation of an image into something that
is more meaningful and easier to analyze. In the segmentation-based OCR
approach, the feature extraction and recognition process depend on the qual-
ity of segmented parts [Nag00]. Therefore, in this approach, segmentation is
the fundamental step for designing an efficient OCR system. However, the
following challenges have been considered as the major bottleneck of seg-
mentation task in document image analysis [CL96, Eik93].
• Over segmentation problem: Due the nature/complexity of the writ-

ing system, touching and fragmented characters being interpreted as
one single character, or a part of a character is believed to be an entire
symbol.
• Document scanning problem: During scanning, joints will occur if the

document is a dark photocopy or if it is scanned at a low resolution. The
characters may be split if the document is scanned at a high resolution
or digitized using a light photocopy.
• Ambiguity of noise and/or texts: Understanding dots and diacritics as

noise wrongly for and vice versa. As a result, a wrong data or non-
texts being sent for recognition. This is mainly happens if characters
are geometrically connected to some graphics.
Segmentation-based OCR techniques have been used for years, during this
period significant works have been done for recognition of different scripts
[PC04, CL96, LV12] and it has been also widely applied for Amharic script
recognition [Mes08, AB08, MJ07, CH03] as well. The smallest individual
components feed for the final recognition stage, in the segmentation-based
OCR technique, are characters. In general, character segmentation is achieved
using either explicit or implicit segmentation techniques, where the former
segmentation techniques are employed in segmentation-based OCR while the
later segmentation technique is the feature of segmentation-free OCR ap-
proach [Cho14, HS95, SG97]. The detailed procedures of this segmentation
technique are described as follows:
• Explicit segmentation: It is a pure segmentation technique in which

1. Classical OCR Approach 15
the sequence of characters in the input document image is segmented

into sub-images of individual characters and is then fed to the charac-
ter recognizer for classification. The vertical segmentation approach
lies in the category of explicit segmentation. In this approach, after the
preprocessing stage of the input handwritten/printed word image, the
word image is scanned from top to bottom [CRA13]. Algorithms in
this segmentation technique follow a cut rule which normally takes a
set of heuristics and information of the foreground, background, or a
combination of them into consideration to generate potential segmen-
tation cuts. In the vertical segmentation technique, once the potential
segmented columns of the image are identified by assigning 00 s and
10 s then the cut point of a character is obtained using histogram profile
projection. In explicit segmentation, the over-segmentation problem
occurred either when the two consecutive characters in the word im-
age are connected by a ligature or when characters are open characters
[CRA13, RMS09].
• Implicit segmentation: It is commonly called recognition-based seg-

mentation and it uses each pixel column as a potential cut location. In
the implicit segmentation approach both segmentation and recognition
of characters are achieved at the same time [CdSBJB+ 06, RMS09]. The
implicit segmentation approach split words into segments which should
be characters, and then pass each segment to a classifier. If the classifi-
cation results are not satisfactory, call segmentation once more with the
feedback information about rejecting the previous result. The implicit
segmentation approach provides all the tentative segments and lets the
recognizer decide the best segmentation hypothesis. However, there is a
trade-off in selecting the number of segments for a word. Computation
becomes efficient if the number of segments is less but wide-character
may be covered in the hypothesis. Whereas, for a large number of seg-
ments, it is computationally expensive.
Like explicit segmentation, implicit segmentation is also exposed for

over-segmentation. To overcome segmentation problems, the implicit
segmentation approach follows a two-stage process. In the first stage,
a heuristic rule-based segmentation scheme is applied to segment the
words. In the second stage, verification is performed. Then segmen-
tation will be finalized based on recognition [RMS09]. Such segmen-
tation techniques are usually applied in the segmentation-free OCR ap-
proach, where CTC network layer is used in the output layer to calculate
መገንባት
መ ዐዐ
Figure 2.1: Character segmentation process and segmentation problems
on sample Amharic text. Character Œ is over segmented into
a a (i.e a single Amharic character pronounced as me is con-
sidered as two other characters pronounced as 0 a).
the loss and decode the feature sequence.
The majority of errors in segmentation-based OCR-systems are often due to

problems in the scanning process and then poor segmentation process result-
ing in joined or broken characters. Confusion between text and graphics or be-
tween text and noise are also the other cause of errors in segmentation-based
OCR systems [Eik93]. The sample segmentation problem in Amharic script is
illustrated in Figure 2.1. The most common type of segmentation-based OCR
model is template matching. In this method, characters are extracted and are
subsequently matched against all the possible templates of the characters. The
recognition is achieved by performing a normalized cross-correlation between
a template character and a new character to be matched [CAC+ 04]. Compar-
ison between characters is made to detect the highest matching area. This
comparison is done by sliding a template character against the new character
by one pixel at a time, from left to right and top to down. The value at each
point, which tells how similar the template is to that particular area, can be
computed using the cross-correlation function f which is the basic similar-
ity measure used in practice and the formulation is given in Equation (2.1)
2. Modern OCR Approach 17
[UH16].
− I¯u,v ][T (x − u, y − v) − T̄ ]
P
x,y [I(x, y)
f (x, y) = qP (2.1)
¯ 2 P x,y [T (x − u, y − v) − T̄ ]2
x,y [I(x, y) − I u,v ]
where I is the image, T is the template, T̄ and I¯u,v is the mean of the template
and mean of the region under the template I(x, y) respectively.
In practice, nowadays, the use of segmentation-based methods is quite lim-

ited as this method greatly suffers from image noise and font variations. In
addition, it is script-dependent and may not work well if we applied it for the
segmentation of texts written in any other script. For example, the technique
developed for segmenting touched characters in Arabic script may not work
well to segment touched characters of texts written in Amharic scripts.
However, the segmentation-based OCR approach has been remained state-

of-the-art for a very long time, in the field of OCR, until it is finally surpassed
by segmentation-free OCR techniques. The next section presents the details
of segmentation-free OCR approaches, which is called the modern OCR ap-
proach in this thesis, and highlights the methodologies employed for script
recognition under this approach.
2.2 Modern OCR Approach
This section describes modern OCR approach, which is commonly called

segmentation-free OCR, in detail. Although segmentation-based OCR ap-
proach widely used in literature, segmentation inaccuracy has been the major
drawback. Lately, the research in the domain of optical character recognition
has moved towards segmentation-free approach that can incorporate contex-
tual information in the recognition process [All95, Nag92]. The modern OCR
approaches perform the character localization and recognition at a text-line
level; thereby avoiding the need of explicit word or character segmentation.
Recognition in segmentation-free OCR approaches does not only depend on
the features of a single character rather it also depends on the basis of its
surrounding context.
Segmentation-free approaches for OCR, differ from the segmentation-based

methods in the level of segmentation required from layout analysis steps. Tra-
ditionally such approaches need the extraction of words or part of words for
a given text-line. The whole word or part of word are then recognized by
the recognition engine. In segmentation-free OCR approaches, discriminat-

ing features are extracted from the whole word or text-line. These features
are then used to train a classifier to recognize the whole word or complete
text-line. For a long period of time, Hidden Markov Model (HMM) were the
state-of-the-art and preferred method [MF06, MSLB98] until the introduction
of deep neural networks and it was surpassed by Recurrent Neural Networks
(RNNs)-based OCR methods.
2.3 Deep Neural Networks
This section describes deep neural networks, an artificial neural network with
multiple layers between the inputs and outputs, in detail. The concept of the
deep neural network is originated from Artificial Neural Networks (ANN)
[HS06], which are inspired by the structure and function of the biological
brain. Until the birth of deep learning [HOT06], ANN and other statisti-
cal classifiers such as Support Vector Machines (SVM), K-Nearest Neigh-
bor (KNN) [HM16], and Random Forests (RN)[Pal05] have been adopted
and widely implemented for various applications. Those classical machine
learning techniques, in general, consists of two modules; the feature extractor
and the classifier module. The feature extractor is responsible for the ex-
traction of discriminant features, usually called handcrafted features, from
the given data. Once those features are extracted, they are subsequently fed
into the classifiers. However, after the introduction of deep neural networks
such as Convolutional Neural Networks (CNNs) such stepwise procedures
are put to one unified framework, where both the feature extractor and clas-
sifier modules are trained in an end-to-end fashion using back-propagation
[LFZ+ 15, SJS13].
2.3.1 Convolutional Neural Networks
Convolutional Neural Networks (CNNs) are a specialized artificial neural net-

work which is designed with the ability to process and extract spatial infor-
mation from image data [LBBH98]. There are various architectures of CNNs
such as, LeNet [LBBH98], AlexNet [KSH12], GoogLeNet[SLJ+ 15], ResNet
[HZRS16], and VGGNet [SZ14], which basically consist of 2 layers in com-
mon namely convolutional layers, and pooling layers. Convolutional layers
convolve around the input image with multiple convolutional filters, usually
performs a dot product between filters and local regions, and pass the output
3. Deep Neural Networks 19
to the next layer, while pooling layers are responsible to reduce the spatial
size of the input [PZLL20].
In addition to these two basic layers, CNNs may consist of fully connected
layers which take the flattened matrix from previous layers and compute the
class score and they may also consist of various network hyper-parameters
such as stride, padding, and depth (number of filters) which could control
the learning algorithms [BSS17]. Convolutional neural networks has been
successfully applied for various computer vision applications such as, im-
age recognition [TT18], facial recognition [HML+ 18], image segmentation
[YNDT18], and character recognition [BHL+ 19b]. CNNs are the dominant
methods in various image-based tasks and they still continue to attract the
interest of the machine learning community across many fields including re-
searchers from document image analysis and recognition.
2.3.2 Recurrent Neural Networks
Recurrent Neural Networks (RNNs) are designed to address the temporal

behavior of input data, which is not easily handled by convolutional neu-
ral networks. The intuition of using RNNs for sequential data is the ability
of recurrent networks to learn and remember over long sequences of inputs.
In sequence-based tasks such as handwritten recognition [LGF+ 07], speech
processing [GMH13], and machine translation [LPM15], we usually need to
know what was in the past and what is coming next to better understand and
detect the present context.
There are various architectures of recurrent neural networks ranging from a
typical vanilla recurrent network, which usually suffers from a problem called
exploding and vanishing gradient [RTAS20], to advanced RNN architectures
such as Long Short-term Memory (LSTM) and Gated Recurrent Unit (GRU)
network, that use gates to control information flow and remedy the gradi-
ent problem of traditional RNN. Recurrent networks are trained by back-
propagation though time [DJH01, Wer90], where the loss (l) is computed
based on the loss at every time step as,
T
X
l= l(ŷi , yi ) (2.2)
i=1
where T , ŷ, y denotes the maximum time-step, prediction and ground-truth.

For the input variable x at time step t, xt , the output of the RNN layer at
time-step t can be computed as,
Ot = ht Wo + bo (2.3)
where Ot is the output, Wo is the weight at the output layer and bo bias and ht
is the hidden state variable at time-step t which can be expressed as follows,
ht = f (xt Wxh + ht−1 Wh + bh ) (2.4)
where f is an activation function, Wxh and Wh weights at current time-step(t)

and previous time-step (t − 1), ht−1 is the hidden state at (t − 1) and bh is the
bias term.
To process data both spatially and temporally, researchers recommended
using a hybrid of CNNs and RNNs. This combination has been successfully
applied in many domains including image captioning and visual question an-
swering [WSW+ 17], handwritten recognition [ZDD18]. The next section
presents the two state-of-the-art techniques for sequence-to-sequence tasks
from the perspective of optical character recognition.
2.4 Sequence-to-sequence Learning for OCR
A more recent approach in the segmentation-free OCR domain is attention-

based sequence-to-sequence learning. However, the use RNNs in combina-
tion with CTC approach are also widely applied to date. These two sequence-
to-sequence learning approaches are the main focus, of this thesis, and the
details are discussed in the next sections. Section 2.4.1, presents the details
of CTC-based sequence learning and attention-based sequence learning is de-
scribed in section 2.4.2.
2.4.1 CTC-based Sequence Learning
Connectionist Temporal Classification (CTC), introduced by Graves [GFGS06],

is a type of neural network output and associated scoring function, for train-
ing RNNs such as Long Short-Term Memory (LSTM) networks to tackle se-
quence alignment problems where the timing is variable. CTC-based super-
vised training of RNNs, on sequential data, has shown great success in many
machine learning areas including end-to-end speech recognition [ZYDS17]
and handwritten character recognition [LGF+ 07].
4. Sequence-to-sequence Learning for OCR 21
In CTC-based sequence learning, the input is a sequence of observations,

and the output is a sequence of labels, which can include a blank token as an
additional outputs [GFGS06]. For a given training data D, the CTC objective
function minimizes the negative logarithm of likelihood loss formulated as
Equation (2.5). X
lCT C = −log p(π/x) (2.5)
π
where x = x1 , x2 , ..., xT is the input sequence with length T , z = z 1 , z 2 , ..., z C

is the corresponding target for C < T and (x, z) ∈ D. p(πt /x) is computed
by multiplying the probability of labels along the path π that contains output
label over all time steps t as shown in Equation (2.6).
T
Y
p(π/x) = p(π t , t/x), (2.6)
t
where t is the time step and π t is the label of path π at t. CTC is independent
of the underlying neural network structure rather it refers to the output and
scoring. The difficulty of training comes from there being many more obser-
vations than the actual labels. For example in a text-line, there can be multiple
time slices of a single character. Since we don’t know the alignment of the
observed sequence with the target labels, we predict a probability distribu-
tion at each time step [Han17] and then the CTC is used to align the output
activation of a neural network with target labels.
A target label in path π is obtained by mapping reduction function B that

converts a sequence of Soft-max output for each frame to a label sequence,
which will be less than or equal to the length of the frame, by first remov-
ing repeated characters and then blank (φ) tokens of the given sequences.
As an example, for a give Eighteen sequence length observation (o), o =
φaaφmmφhφaaφrrφicφ. Then, the paths are mapped to a label sequence
as ls = B(o) = B(φaaφmmφhφaaφrrφicφ) = B(φaφmφhφaφrφicφ) =
’amharic’, where B is a reduction mapping function which works by first col-
lapse the repeated tokens and then remove blanks.
The target sequence probability y from input sequence x is the sum of the
probability of all paths by reducing each path to this label sequence using B
and this probability is again obtained using Equation (2.7).
X
p(y/x) = p(π/x), (2.7)
π∈B(y)
Predicted text
አማርኛ የኢትዮጵያ ፡ መደበኛ ፡ ቋንቋ ፡ ነው::
CTC
RNN
RNN
.. Output-layers
Input-layers
.
RNN
RNN
Hidden-layers
አማርኛ የኢትዮጵያ ፡ መደበኛ ፡ ቋንቋ ፡ ነው:: Ground-truth text
አማርኛ የኢትዮጵያ ፡ መደበኛ ፡ ቋንቋ ፡ ነው:: Input Image
Figure 2.2: Illustration of CTC-based sequence learning. In CTC-based

sequence learning paradigm, the CNNs and RNNs are trained to-
gether in a supervised manner using the whole text-line image and
the corresponding ground-truth sequence. The output activations
are aligned with the ground-truth using CTC algorithm. Then the
CTC network is responsible for error calculation and transcrip-
tion.
Figure 2.2 shows general flow diagram of CTC-based sequence learning

paradigm, that uses Bi-LSTM as a sequence learner and CTC cost of target
transcription, for an input sequence of sample Amharic text-line image.
2.4.2 Attention-based Sequence Learning
Attention has been used in different contexts across multiple disciplines, for
example, psychologists define Attention as the behavioral and cognitive pro-
cess of selectively concentrating on a discrete aspect of information while
ignoring other perceivable information [Jam07]. Human visual attention, Bi-
ologically, allows us to focus on a certain region with high resolution while
perceiving the surrounding in low resolution and then adjust the focal point
or do the inference accordingly [All89]. In Computing, Attention is one of
the components of the network architecture and is in charge of managing and
quantifying the interdependence of sequences [BCB14].
Attention is, to some extent, motivated by how we pay visual attention to

different regions of an image or correlate words in a sentence. It is first intro-
duced, by Bahdanau [BCB14], as a solution to the limitation of the encoder-

decoder model in sequence transcription. The architecture of encoder-decoder
networks has two modules, the encoder, and the decoder. Basically, both the
encoder and the decoder modules can be any variant of the neural network,
but they usually use recurrent neural networks, where the role of the encoder
module is to transform the given variable-length input sequence into a fixed-
length sequence vector and then the decoder module uses these vector so as
to generate the target output sequences.
To generate the output sequence from a given input sequence, the encoder
and decoder modules are unified and process the information from end-to-
end. All the information of the input sequence goes to the decoder module
through the last hidden state of the encoder module which usually causes
a bottleneck on this hidden state. It was a challenge for encoder-decoder
models until the introduction of the attention mechanism, which advances
the working principle of the encoder-decoder model and makes it the most
competitive model in sequence-to-sequence problems to date.
Attention-based network models have emerged as an extension of the exist-
ing encoder-decoder model so as to solve the problems that occur during de-
coding long sequences. Nowadays, various attention models are proposed for
solving various sequence-to-sequence tasks. This section introduces the two
most commonly used attention models, based on work of Bahdanau [BCB14],
Xu [XBK+ 15], and Luong [LPM15], namely Soft/global and Hard/Local at-
tention respectively. The differences, in these attention models, lie mainly in
their system architectures, scoring functions, and computations. The details
of these commonly used attention mechanisms are discussed as follows.
• Soft Attention: This attention considers the entire input state spaces
of the encoder to drive the context vector and it is usually referred
to as Global attention [XBK+ 15, LPM15] and/or Additive attention
[BCB14]. As explained in the work of Bahdanau [BCB14], the in-
tuition of implementing attention was with the aim of improving the
sequence-to-sequence model in machine translation by aligning the de-
coder with the relevant input sentences. The soft-attention mechanism
uses Concat, as an alignment scoring function, whereby the decoder
hidden state is added to the encoder hidden states. Since soft/global
attention attends all the encoder hidden-states, it is computationally ex-
pensive, especially for longer input sequences.
• Hard Attention: The second type of attention which is proposed to
address the limitations of global [LPM15] or soft [XBK+ 15] attention
and it is often referred to as Multiplicative attention or Local attention

[LPM15]. Unlike the soft attention, hard attention attends part of an
input sequence. The local attention proposed by [LPM15] consists a
blended features of both soft and hard attentions of [XBK+ 15] and,
therefore, it avoids the expensive computations experienced in atten-
tion model of [BCB14, XBK+ 15]. In addition to the alignment score
function, Concat used in Additive attention, this attention mechanism
[LPM15] introduces two other alignment score functions the Dot and
General. The Dot alignment score function produces the alignment
score by multiplying the hidden states of the encoder and the hidden
state of the decoder. Except the weight matrix added in the formula-
tion, the General alignment score function is similar to the Dot align-
ment function.
As stated in the literature, the way that the alignment score is calculated and
the position at which the attention mechanism is being introduced in the de-
coder are the major differences in various attention mechanisms. For example,
in Luong’s local attention, Concat alignment score function, the decoder hid-
den state and encoder hidden state will not have their own individual weight
matrix, but a shared one instead, unlike in the attention model proposed by
Bahdanau. Therefore, the decoder hidden state and encoder hidden states are
added together first before being passed through a linear layer, and then, after
being passed through the Linear layer, a tanh activation function will be ap-
plied to the output before being multiplied by a weight matrix to produce the
alignment score. A generic flow of Attention-based sequence learning, which
is designed based on the work of Bahdanau [BCB14], is shown in Figure 2.3.
The sequence-to-sequence learning was born in the field of language mod-

eling [SVL14] that takes a sequence of inputs and then outputs another se-
quence of items where both sequences can be of arbitrary lengths. The encoder-
decoder architecture for recurrent neural networks is proving to be powerful
on various sequence-to-sequence prediction problems such as machine trans-
lation and caption generation. However, a fixed-length context vector design
of this model is a critical disadvantage. It has forgotten the first part once it
completes processing the whole input specifically on long input sequences.
Therefore, the attention mechanism was originally introduced to help mem-

orize long source sentences in Neural Machine Translation (NMT) [BCB14].
Rather than building a single context vector out of the encoder’s last hidden
state, the advantage of Attention is to create shortcuts between the context
vector and the entire source inputs. The weights of these shortcut connections
y0 y1 y2 yTy
..
.
Softmax Softmax Softmax Softmax
..
.
s1 s2 sTy-1
Decoder-RNN Decoder-RNN Decoder-RNN
..
.
Concat Concat Concat
..
.
s0
context1 context2 contextTy
attention attention attention
..
.
h1 h2 hTx
h0 Encoder-RNN Encoder-RNN Encoder-RNN
..
.
X1 X2 XTx
..
Input-layers
.
Figure 2.3: Illustration of Attention-based sequence learning. In attention-
based sequence learning paradigm, the two RNNs (Encoder and
Decoders) are employed. An encoder processes the input se-
quence (x1 , ..., xTx ) with length Tx and compresses the informa-
tion into a context vector (contexti ) of a fixed length. This repre-
sentation is expected to be a good summary of the meaning of the
whole input sequence while the decoder is initialized with the con-
text vector to emit the transformed output sequence ( y1 , ..., yTy )
of length Ty .
are customizable for each output element. Since the context vector has access
to the entire input sequences, we don’t need to worry about forgetting. The
alignment between the source and the target is learned and controlled by the
context vector.
Attention-based sequence-to-sequence learning employees two recurrent

networks; the Encoder-RNN encodes the input sequence and outputs a se-
quence of the same length vector while the Decoder-RNN interprets the en-
coded fixed length vector and generates a target sequences. Let’s define a
generic Attention-based encoder-decoder network model, which is introduced
for solving a sequence-to-sequence tasks specifically focus on NMT. For an
input sequence, X =[x1 , x2 , ..., xT x ], of length T and an output target se-
quence, Y =[y 1 , y 2 , .., yT y ], with different length of t. Suppose the encoder is
X X X
w1 w2 ... wTx
SoftMax
a0,1 a0,2 a0,Tx
tanh tanh ... tanh
S0
...
h1 h2 hTx
Figure 2.4: A typical attention model. This design computes the output c
from the initial decoder hidden state s0 and the part of the given
sequence hi , where the decoder hidden states at t − 1 are first
concatenated with each encoder hidden states and passes through
tanh activation. The tanh layer can be replaced by any other net-
work that is capable to produce an output form s0 and hi . Here,
wi and ai denotes attention weights and alignment scores respec-
tively.
→
−
a bidirectional RNN with a forward hidden states h i and a backward hidden
←−
states h i . With the motivation to include both the preceding and following
words in the annotation of one word, the encoder state hi is the concatenation
→
− ← −
of the two hidden states and represented as, hi =[ h i , h i ] for i = 1, .., T .
The decoder network has hidden state sj = f (sj-1 , y j-1 , cj ) for the output
sequence at position j, j = 1, ..., t, where the context vector cj = Ti=1
P x
sji hi ,
is a sum of hidden states of the input sequence, weighted by alignment scores,
aji = Align(y j , xi ), which shows how well two sequences y j and xi are aligned.
The alignment score aji , as described in the work of Bahdanau [BCB14], is
computed by a feed-forward network with tanh activation function which
is jointly trained with other parts of the model. Then the alignment score is
given as, score(sj , hi )= vaT tanh(Wa [sj ; hi ]), where both v a and W a are weight
matrices to be learned in the alignment model. A typical attention layer ar-
chitecture that computes a context vector for the first decoding time step is
5. Chapter Summary 27
depicted in Figure 2.4. Even though the Attention mechanism is designed and
has been used to solve problems in NMT, nowadays, it is also widely applied
in various fields such as image captioning [LTD+ 17], visual question-answer
[AHB+ 18], and OCR [ZDD18, PV17, CV18] and promising experimental
results are reported.
2.5 Chapter Summary
Trends in OCR research and widely used character recognition methodolo-

gies are highlighted in this chapter. Research in the OCR model development
can be either segmentation-based or segmentation-free. Since segmentation-
based OCR depends on the segmentation stage which is usually affected by
many factors, (like the nature of the scripts and quality of documents), the
segmentation-free approach is the choice and state-of-the-art techniques used
to develop the OCR model. The segmentation-free OCR approach can be
either a CTC-based or attention-based sequence learning. The CTC-based
networks have been used to tackle sequence-to-sequence problems where the
timing is variable. In this approach, the current output of the CTC model
relies on the current input and does not account for context weighted inputs
and it assumes that each output is conditionally independent of the other out-
puts. Therefore, attention-based sequence-to-sequence learning has revolu-
tionized and considers a standard fixture in most state-of-the-art NLP models.
Attention-based sequence learning focuses on specific sequences and con-
siders context-weighted inputs when formulating a response, unlike in CTC-
based sequence learning. Nowadays, attention-based networks are widely
used in solving problems in the field of OCR. The research attempts made
to develop Amharic OCR model so far are segmentation-based and there is
no workable OCR of this language. Thus, in this thesis, such a research
gap is further investigated by employing state-of-the-art, segmentation-free,
sequence-to-sequence learning techniques. In addition to solving segmentation-
related challenges, the existing sequence-to-sequence approaches are exten-
sively investigated and then the best features from the CTC and attention-
based techniques are integrated for an improved feature representation and
output sequence generation.
28
Chapter 3
Amharic Script and Writing System
This chapter provides an overview of the script, Amharic, which we are go-
ing to deal with in this thesis. Amharic is an exotic script that has its own
indigenous alphabet and a rich collection of documents. Understanding the
language and its writing system are indispensable steps in Natural Language
Processing (NLP) tasks in general. OCR application heavily depends on the
writing system of the language. This chapter first presents the demographic
distribution of Amharic speakers and the genetic structure of Amharic script.
Furthermore, the writing system and orthographic identity of Amharic char-
acters with their corresponding unique features, in the script, have been pre-
sented. This chapter also provides a brief description of related works, on
Amharic script, which has been done prior and/or in parallel with this thesis.
Finally, the summary of the chapter is presented.
3.1 Amharic Script and its Genetic Structure
In Ethiopia, there are more than 86 languages, that are divided into Semitic,
Cushitic, Omotic, and Nilo-Saharan groups [EF20], with up to 200 differ-
ent dialects spoken [MJ05, Mes08]. The languages are written with either
Ge’ez/Amharic and/or Latin scripts. Of the 86 languages in Ethiopia, 41 are
living languages at the institutional level, 14 are developing, 18 are vigor-
ous, 8 are in danger of extinction, and 5 are near extinction [EF20]. Of these
languages, Amharic is an official working language of the Federal Demo-
cratic Republic of Ethiopia and is spoken by more than 50 million people
as their mother language and over 100 million as a second language in the
29
30 Amharic Script and Writing System
Figure 3.1: Demographic distribution of Amharic speakers. Amharic is

widely spoken in different countries while a country marked with
yellow color is a nation which uses Amharic as an official working
language3 .
country [pre, MM18]. The demographic distribution of Amharic speakers

is illustrated in Figure 3.1. In addition to Ethiopia, it is also widely spoken
in other countries like Eritrea, USA, Israel, Sweden, Somalia, and Djibouti
[MJ07, Mey06].
The script of Amharic language is originated from the Ge’ez script, an an-
cient Ethiopian script, around 300 A.D. Amharic script consists of the full set
of Ge’ez characters and includes its own additional characters. Nowadays,
Amharic characters are used to write various languages in Ethiopia includ-
ing Amharic, Guragegna, Awi, Argoba, and Tigrigna. Figure 3.2 depicts the
3
https://www.adams.africa/works/ethiopia/
2. Orthographic Identity of Amharic Characters 31
Afro-Asiatic
Berber Egyptian Semitic Chadic Cushitic Omotic
East-Semitic West-Semitic
Central-Semitic Ethiopian-Semitic
North South
Ge'ez Tigrigna Gurage Amharic
4
Figure 3.2: Family of Amharic language.
genetic structure of Amharic language. Amharic script is designed as a mean-

ingful and graphic representation of African knowledge systems which is con-
sidered as one of the signal contributions made by Africans to world history
and cultures. It is created to holistically symbolize and locate the cultural and
historical parameters of the Ethiopian people [Asn12].
3.2 Orthographic Identity of Amharic

Characters
Writing systems can be placed into different categories such as alphabetic,

syllabic, or logographic [PC00, Mes08]. In addition, any particular writ-
ing system may have also attributes of more than one category [Cou03]. In
the alphabetic category, a standard set of letters formed from consonant and
vowel combinations represent speech sounds (e.g. Latin alphabet). In a syl-
labic/Abugida, each symbol correlates to a syllable (consonant-vowel pairing)
in which consonants are written as the main letters, and then special sym-
bols/diacritics are used to indicate which vowels follow the consonant (e.g.
4
Obeid, Nassim. ”Etymological study of semitic languages.” (2009).
Amharic and Devanagari characters). In a logography, each character repre-

sents a semantic unit such as a word or morpheme (e.g Chinese character).
Generally, alphabetic scripting typically uses a set of fewer than 100 sym-
bols to fully express a language, whereas syllabic can have several hundred,
and logographies can have thousands of symbols [Cou03]. Many writing sys-
tems also include a special set of symbols known as punctuation which is used
to aid interpretation and help capture variations in the message’s meaning
that are communicated verbally by cues in timing, tone, accent, or inflection
[Aro85].
Amharic uses the writing system called Fidel, which grew out of the Geez
abugida, and it consists of about 317 different characters which are written and
read, as in English, from left to right in horizontal lines [BHM+ 20, MJ07].
In Amharic writing system, each symbol represents a syllable consisting of
combinations of consonants and vowels as a single alphabet. The basic signs
are modified in a number of different ways to indicate the various vowels from
their corresponding consonants [BHL+ 19b].
All vowels and labialized characters in Amharic script are derived, with
a small change, from the 34 consonant characters. The changes involve a
considerable regularity for modifying the structure of consonant characters
by adding a straight line, or shortening and/or elongating one of its main legs.
It also includes the addition of small diacritics, such as strokes or loops/circles
to the right, left, top, or bottom of the character. The 34 consonant Amharic
characters and their corresponding six vowel variants of each character are
illustrated in Figure 3.3.
In addition to the 238 basic characters, Amharic script uses nearly 79 other
characters which are depicted in Figure 3.4 includes 50 labialized characters,
9 punctuation marks, and 20 numerals [BHM+ 20, HA13]. This brings the
full sets of Amharic Fidel used in the Script to 317 characters. A matrix that
consists of all these Amharic characters is called Fidel-Gebeta Unlike vowels
which are derived with a regular pattern modification of consonant Amharic
characters, Labialized Amharic characters are derived with less regular pat-
terns modification. The modern Amharic writing system particularly uses
few of Amharic punctuation marks like full stop ( ~ ), Comma (,) and preface
colon ( { ) [Ale97]. Others have largely fallen into disuse and a number of
punctuation marks, such as exclamation mark ! and the question mark ?, have
been adopted from the Western practices [Tes10]. Amharic numbering sys-
tem represents numbers from one to ten, multiples of ten (twenty to ninety),
hundred, and thousand. However, Amharic numbering system has long since
2. Orthographic Identity of Amharic Characters 33
Figure 3.3: List of basic Amharic characters. Row-column-wise (34 × 7)

order, where half of the characters are give on the left and the
other halves are given on the right side of this figure. writing and
reading are from left to right: The first column gives the consonant
characters while the next six columns are derived by combining
the consonant with vowel sounds. Across all the seven orders, the
shapes are constrained by the vowels they represent.
been retired to a reserved use primarily for calendar dates and sometimes for
paging. As stated, in literature, Amharic writing system does not have a sym-
bol for zero, negative, decimal point, and mathematical operators. Therefore,
the Arabic numerals are used everywhere else while Latin mathematical op-
erators are used for arithmetic computation.
3.2.1 Features of Amharic Characters
Since a character in Amharic script is formed from its corresponding conso-

nant by adding diacritics, most of the characters have some common notable
characteristics such as shape similarity, vowel formation, structural relation-
ship, height, and width difference. Each symbol is written according to the
sound that goes with it. The vowels are integrated into the characters by mod-
ifying the base characters in some form, which together represent syllable
combinations consisting of a consonant and vowel. Thus Amharic writing
system is often called syllabic rather than alphabetic. Some of the features of
Amharic characters could be described as
• The direction of writing and reading the characters is from left to right
and top to bottom, where characters are written in a disconnected man-
ner and proportional spacing.
ቈ kʷ̣ ä ቊkʷ̣ i ቋkʷ̣ a ቌkʷ̣ e ቍkʷ̣ ə

ኈḫʷä ኊḫʷi ኋḫʷa ኌḫʷe ኍḫʷə
ኰkʷä ኲkʷi ኳkʷa ኴkʷe ኵkʷə
ጐgʷä ጒgʷi ጓgʷa ጔgʷe ጕgʷə
ሏlwa ሗhwa ሟmwa ሧszwa ሯrwa
ሷswa ሿshwa ቇqwa ቧbwa ቯvwa
ቷtwa ቿcwa ኇxwa ኗnwa ኟñwa
ኧʼwa ኯkwa ዟzwa ዧzhwa ዯywa
ዷdwa ዿphwa ጇjwa ጏ gwa ጧthwa
ፏfwa ጯchwa ጿtswa ፇtzwa ፗpwa
a
፩1 ፪2 ፫3 ፬4 ፭5 ፠ section mark ፡ word separator

፮6 ፯7 ፰8 ፱9 ፲10 ። full stop ፣ comma
፤ semicolon ፥ colon
፳20 ፴30 ፵40 ፶50 ፷60
፦ Preface colon ፧ question mark
፸70 ፹80 ፺90 ፻100 ፼1000 ፨ paragraph separator
b c
Figure 3.4: Other symbols of Amharic Fidel. (a) Labialized characters. (b)
Numerals and (c) Punctuation marks. In most moder Amharic
documents, Amharic numerals are replaced by Latin numbers
while some of the punctuation marks such as section mark and
paragraph separators are not used any more.
• There is a space between words which usually either a blank space or

an Ethiopic two dots/word separator( : ). Sometimes, in early printed
books, a word separator followed by blank space can be used together
(see Figure 3.5).
• Vowel characters are derived from the consonants, with a small modi-
fication of their shapes. For example, as shown in Figure 3.3, the 1st
column consists of the basic symbol of Amharic characters with no ex-
plicit vowel indicator which is usually called consonants. In the 2nd
and 3rd columns the corresponding consonants are modified by adding
a hyphen (-) like diacritics in half-way down the right leg and the base
of the right leg of the character respectively. The 4th column has a short
3. Challenges for the OCR of Amharic Script 35
left leg while the 5th column has a loop/circle on the right leg. The
6th and 7th columns are less structured and lack consistency, but some
regular modification has been made on either the left or right leg of the
base characters. Otherwise, some other modifications such as loops or
small diacritics could also be added to the top of the characters.
• Sentences are separated with Ethiopic four dot/full-stop ( ~ ) while the
paragraphs are separated with a full line of horizontal space or can be
started with indentation.
• Unlike the Latin characters, there is no capital and lower case distinc-
tion among Amharic characters.
• There is a noticeable variance in width and height among characters,
some of them such as character ˜, ˜ and ¥ are longer than
character Œ, P and Â vertically, while the later characters are
longer than the former characters horizontally.
In addition, the line of printed Amharic script lies at the same level, having
no ascent and descent features which always have a white line between two
consecutive lines of characters. Sample Amharic scripts, including some of
the features stated above, are illustrated in Figure 3.5.
3.3 Challenges for the OCR of Amharic Script
Even though there is a great demand for Amharic OCR, research in this field
is still behind that of the well-known scripts, such as Latin, and even it is the
untouched one. Amharic script is much more complex as compared to Latin
scripts. The character shapes in Amharic script are far too many and vary
widely with different non-standard fonts. Since Amharic is an indigenous
language with its own unique characters and writing system, understanding
the characteristics and distinctive features of this script is challenging and
also interesting because it has an enormous use for the development of optical
character recognition system for Amharic script. Some of the generic and
script/language-specific challenges are outlined below:
• Visual similarity among characters: This is due to shape similarity
between consonant characters and also across vowels which are derived
from the base characters with small modifications. Since characters in
Amharic script are derived from 34 basic characters, there is a high
5
https://www.bbc.com/amharic/news-51780956
የጤና:ተቋማት:እንደሚሉት:በሽታውን:ለመከላከል:እጅን:በመደበኝነት:በደንብ:መታጠብ:እጅግ:ጠቃሚ:ነው። እስካሁን:
በሽታው:እንዴት:እንደሚተላለፍ:በእርግጠኝነት:የሚታወቅ:ነገር:የለም። ነገር:ግን:የኮሮና:አይነት:ቫይረሶች:በበሽታው:የተያዙ
ሰዎች:በሚያስሉበትና:በሚያስነጥሱበት:ወቅት:በሚያወጧቸው:ጥቃቅን:ጠብታዎች:አማካይነት:ነው። ስለዚህም:ስናስልና:
ስናስነጥስ:አፍንጫና:አፋችንን:በሶፍት:ወይም:በክርናችን:እንሸፍን፤ ፊታችንን:እጃችንን:ሳንታጠብ:አንንካ። በተጨማሪም:
ከታመሙ:ሰዎች:ጋር:የቀረበ:ንክኪ:ከማድረግ:መቆጠብ:በሽታውን:ለመከላከል:ያግዛል። የህክምና:ባለሙያዎች:እንደሚሉት:
የፊት:መሸፈኛ:ጭንብሎች:የሚፈለገውን:ያህል በሽታውን:ለመከላከል:አያስችሉም። በበሽታው:ተይዣለሁ:
ብለው: የሚያስቡ:ሰዎች:አስፈላጊውን:ምክርና:የህክምና:ድጋፍ:ለማግኘት:ለሚመለከታቸው:የጤና:ተቋማት:ማሳወቅ
አለባቸው። በበሽታው:ተያዞ:ይሆናል:ተብሎ:ከሚታሰብ:ሰው:ጋር:የቀረበ:ንክኪ:ከነበረዎት፤ እራስዎን:ለተወሰነ:ጊዜ:ከሌሎች:
ሰዎች:አግልለው:እንዲቆ:ሊነገርዎት:ይችላል። ስለዚህ:በበሽታው:ከተያዙ:ሰዎች:ጋር:የሚኖርን:የቀረበ:ንክኪ:ማስወገድ:
ይኖርብዎታል። በሽታው:ከተከሰተባቸውና:ከሌሎች:አገራት:የተመለሱ:ከሆነ:ምናልባትም:ለተወሰነ:ጊዜ:እራስዎን:አግልለው
መቆየት:ሊያስፈልግዎት:ይችላል።
(a)
የጤና: ተቋማት: እንደሚሉት: በሽታውን: ለመከላከል: እጅን: የጤና ተቋማት እንደሚሉት በሽታውን ለመከላከል እጅን
በመደበኝነት: በደንብ: መታጠብ: እጅግ: ጠቃሚ: በመደበኝነት በደንብ መታጠብ እጅግ ጠቃሚ ነው። እስካሁን
ነው። እስካሁን: በሽታው: እንዴት: እንደሚተላለፍ: በሽታው እንዴት እንደሚተላለፍ በእርግጠኝነት የሚታወቅ
በእርግጠኝነት: የሚታወቅ: ነገር: የለም። ነገር: ግን: የኮሮና: ነገር የለም። ነገር ግን የኮሮና አይነት ቫይረሶች በበሽታው
አይነት: ቫይረሶች: በበሽታው: የተያዙ: ሰዎች: የተያዙ ሰዎች በሚያስሉበትና በሚያስነጥሱበት ወቅት
በሚያስሉበትና: በሚያስነጥሱበት: ወቅት: በሚያወጧቸው: በሚያወጧቸው ጥቃቅን ጠብታዎች አማካይነት
ጥቃቅን: ጠብታዎች: አማካይነት: ነው። ስለዚህም: ስናስልና ነው። ስለዚህም ስናስልና ስናስነጥስ አፍንጫና አፋችንን
ስናስነጥስ: አፍንጫና: አፋችንን: በሶፍት: ወይም: በክርናችን: በሶፍት ወይም በክርናችን እንሸፍን፤ ፊታችንን እጃችንን
እንሸፍን፤ ፊታችንን: እጃችንን: ሳንታጠብ: አንንካ። ሳንታጠብ አንንካ። በተጨማሪም ከታመሙ ሰዎች ጋር
በተጨማሪም: ከታመሙ: ሰዎች: ጋር: የቀረበ: ንክኪ: የቀረበ ንክኪ ከማድረግ መቆጠብ በሽታውን ለመከላከል
ከማድረግ: መቆጠብ: በሽታውን: ለመከላከል: ያግዛል። ያግዛል።
(b) (c)
Figure 3.5: Sample Amharic scripts. (a) Word are separated with two dots.
(b) Words are separated by two dots followed by blank space. (c)
Words separated with blank space. The texts are obtained from
BBC Amharic news5 .
chance of being visually similar; thus the task of recognition becomes

hard for machines as well as humans.
• Character formation inconsistency across columns: Even though,

there is considerable shape modification regularity of characters across
some columns, some of them are more consistent than others while
others are completely inconsistent. Some of them, such as the 3rd and
5th column are more consistent than others, such as the 2nd and 4th
columns, while others, such as the 6th and 7th columns are completely
inconsistent. The 6th column is the least consistent, with the greatest
number of patterns, so that the character formation is largely unpre-
3. Challenges for the OCR of Amharic Script 37
dictable although in the entire Amharic Fidel some clusters of Amharic

characters follow similar patterns modification to the 6th column.
• Printing quality and font variations: Printed documents are also writ-
ten using various fonts, styles, and sizes with different quality of print-
ing materials such as ink, printing paper, and even the printing device it
self. As a result of these features, the shape and appearance of char-
acters vary substantially. For printing in a specific Amharic script,
there are different fonts (for example, Power Geez, Visual Geez, Nyala,
and Agafari) with several stylistic variants such as Normal, Bold, and
Italic. In addition, different font sizes including 10, 12, 14, 16 pixel
[Mes08]. Since these fonts, styles, and sizes produce texts that greatly
vary in their appearances, it is a challenging task for developing charac-
ter recognition systems. However, these challenges are independent of
the scripted nature rather they are incurred due to either font variation
or the printing quality.
• Language-specific challenges: Since indigenous languages of Africans

are underrepresented in the area of Natural Language Processing (NLP),
most of them are resource-limited. As a result, they face many ad-
ditional challenges including a lack of standard representation for the
fonts and encoding, even operating systems, browsers, and keyboards
may not directly support some of these languages. This leads to lan-
guage processing routines than the actual effort required for the OCR
model development.
Even though numerous attempts have been made, to overcome the above-
stated general and language-specific challenges of Amharic script, only few
attempts are made to develop an OCR model for Amharic character recogni-
tion. The OCR models developed by researchers so far are based on classical
machine learning techniques and limited Amharic characters were considered
for training as a result, the OCR model yields poor character recognition per-
formance and even no organized and publicly available OCR dataset. There-
fore, new paradigms should be researched and even the accuracy of the ex-
isting algorithm should be verified and optimized. The next section presents
efforts that have been made for developing Amharic OCR.
3.4 Related Works
Following the 1997 first attempt made by Worku [Ale97], different statis-
tical machine learning techniques have been explored by many researchers
for Amharic script recognition. In 1999, Dereje [Tef99] has conducted re-
search so as to recognize a typewritten Amharic text with the aim of improv-
ing Amharic OCR and recommended as the algorithms adopted for Amharic
OCR should not be very sensitive to the features of the writing styles of char-
acters. A year later, Million [Mil00] adopted a recognition algorithm that can
handle the different typefaces of Amharic characters with the aim to investi-
gate and extract the attributes of Amharic characters so as to generalize the
previous research works on Amharic OCR.
As a continuation of the research activities done so far, Yaregal [Ass02] pro-

posed an integrated approach and come up with a versatile algorithm that is
independent of the fonts and other quantitative parameters of Amharic charac-
ters. In addition, other recent attempts have also been made employing deep
learning techniques [GSTBJ19, RRB18]. However, nearly all of these at-
tempts performed segmentation of text images into words and then into char-
acter level which directly affected the performance of the OCR model.
The only exception was Assabie [AB09], proposed an HMM-based model

for off-line handwritten Amharic word recognition without character segmen-
tation using structural features of characters as building blocks of the recogni-
tion system. Recently published work [ALT18], proposed a segmentation-free
technique based on Long-Short-Term Memory (LSTM) networks together
with CTC (Connectionist Temporal Classification) for Amharic text-image
recognition. However, the datasets are not publicly available and all these
attempts to Amharic OCR neither shown results on a large dataset nor con-
sidering all possible characters.
Although OCR development attempted regarding Amharic script so far is

limited, in this thesis, previous research attempts are taken as a baseline and
further investigations are made so as to facilitate the OCR of Amharic script.
In addition, new Amharic OCR databases are prepared and made freely avail-
able for the public with benchmark results recorded empirically with state-of-
the-art OCR techniques.
3.5 Chapter Summary
In this chapter, we have discussed the overview of Amharic script. The origin
of Amharic script, character sets with its formation, the generic structure, and
its relationship with other Semitic languages are presented. In addition, we
highlight the difficulties and challenges related to Amharic script that have
bearings on the OCR development, especially, the existence of visually sim-
ilar symbols and the use of a number of distinct fonts in printed documents
that are designed in an unstructured manner.
We also present a short survey of related works made for the indigenous lan-
guages with a particular emphasis on Amharic script which is currently used
as the official working language of Ethiopia with its own indigenous alpha-
bets and writing system called Fidel. Amharic script has 34 base characters
each of which occurs in seven orders, which represent syllable combinations
consisting of a consonant and following vowel. The total number of unique
characters, used in Amharic script, is about 317, including 238 basic charac-
ters and other symbols representing labialization, numerals, and punctuation
marks.
Developing the OCR model for an indigenous, but the resource-limited lan-
guage is challenging and new algorithms are required to address these chal-
lenges. Unlike most modern Latin scripts enjoying a collection of resources
including a database and benchmark results of OCR, the scarcity of such in-
need resources makes the OCR development very challenging for Amharic
script. In the next chapter, a baseline database prepared for Amharic OCR
development is described in detail.
Chapter 4
Amharic OCR and Baseline Datasets
This chapter gives a brief explanation about a baseline database developed for
Amharic OCR and techniques employed for Amharic database creation. First,
it gives highlights on the effect of dataset unavailability on Amharic charac-
ter recognition tasks. Next, it introduces related works regarding datasets in
general and in particular for Amharic script. Finally, it gives a summary of all
works and contributions regarding Amharic OCR database.
4.1 Introduction
The shortage of training dataset is one of the limiting factors in developing

a reliable system in the area of image processing and pattern recognition,
in general, and for Amharic OCR in particular. Since the performance of
the OCR model is depending on the size and quality of training datasets, we
need to prepare data that is enough for learning and generalization. Manual
annotation of the dataset requires lots of time and effort. Even, it is sometimes
difficult to transcribe the text image into its corresponding ground-truth texts
unless we are an expert of the language.
The following solutions, as the contribution of this thesis, are proposed

to overcome the shortage of annotated datasets with a particular focus on
Amharic script.
• Synthetically generated Amharic database has been introduced. This

database consists of character-level images and text-line-level images
with their corresponding ground-truth texts in two common Amharic
41
42 Amharic OCR and Baseline Datasets
fonts ( Power Ge’ez and Visual Ge’ez fonts).

• Scanned printed Amharic database. This database contains scanned
printed text-line images written in Power Ge’ez font for the OCR of
Amharic script.
All these databases are named Database for Amharic OCR (ADOCR) and it
is, now, freely available at http://www.dfki.uni-kl.de/˜belay/.
The details of these databases are described in section 4.3.
4.2 Dataset for OCR Tasks
In literature, various researchers use different approaches to prepare a dataset

for developing the OCR model. Some of them employed synthetic text-image
generator engine so as to render a synthetic image with their corresponding
ground-truth texts while others used manual annotation techniques. The other
approach in developing a database is first creating documents electronically
and then printed out. Finally aligned the scanned version of this document
image with the previous corresponding electronically created texts.
Nowadays, the use of synthetically generated data is getting popular in the
area of image processing and document image analysis tasks. There are multi-
ple OCR databases created, and reported in the literature, using this technique.
Sabbour and Shafait [SS13] presented a synthetic text-line image database
called Urdu Printed Text Images (UPTI) database. This database consists of
10,063 Urdu text lines images, synthetically generated with the Nastaliq font,
and their corresponding ground-truth text.
Arabic Printed Text Image (APTI) database is also a synthetically gener-
ated Arabic database developed by Slimane et al [SIK+ 09]. This database is
generated using more than 100 thousand words, 10 Arabic fonts with a variety
of font sizes and styles which are widely used on a computer screen, Arabic
newspapers, books, and many other Arabic documents. The database contains
over 45 million words of images and more than 250 million total characters.
The other type of synthetically generated dataset is the SynthText in the
Wiled Dataset published by Gupta et al [GVZ16]. The database consists of
800 thousand images with approximately 8 million synthetic word instances
that are placed in natural scene images. Each text instance is annotated with its
text word and character level bounding-boxes. There are also other numerous
scanned text images of OCR databases such as IAM-DB Handwritten [MB02]
3. Database for Amharic OCR 43
database, RIMES [GCBG08] database which contains 1539 page of English

texts and 12,723 scanned page of French mails respectively.
A real-world image dataset called the Street View House Numbers (SVHN)
[NWC+ 11] is proposed for machine learning and object recognition tasks.
This dataset consists of 600,000 digits of house-number images with a vari-
able resolution, where 73,257 digits for the training set, 26,032 digits for the
test set, and 531,131 digits of the extra training set with less difficult samples.
SVHN dataset is similar to MNIST [LBBH98] dataset, but SVHN comes from
a significantly harder and unsolved problem which are obtained from house
numbers in Google Street View image. COCO-text [VMN+ 16] is the other
real-world and benchmarking database specifically designed for text detection
and recognition in natural images. The COCO-text dataset consists of 63,000
images and more than 145,000 instances of machine-printed and handwritten
texts.
Chen and Qingcai [ZCW10] proposed the Harbin Institute of Technology
Opening Recognition Corpus for Chinese Characters (HIT OR3C) database
which contains more than 900 thousand images of Chinese characters. The
characters have been collected using a handwriting pad and are recorded
and labeled automatically via the handwriting document collection software
called OR3C Toolkit.
The IMPACT dataset [PPCA13], proposed by Papadopoulos et al, is a
printed text database that contains over 600,000 historic document images
that originate from major European libraries and covering texts from news-
papers, books, pamphlets, and typewritten notes. However, to the best of our
knowledge, there is no publicly available dataset for Amharic script recogni-
tion. In this regard, we have prepared a new public Amharic database and a
baseline recognition system. The dataset is a composition of several datasets,
a character image and a text-line image dataset that includes scanned printed
and synthetic texts.
4.3 Database for Amharic OCR
The primary issue in the field of OCR in general and in particular for Amharic
OCR is a lack of annotated dataset. Therefore, this section describes various
approaches followed to develop the database and the detail of Amharic OCR
database called ADOCR [BHL+ 19a]. The ADOCR database contains char-
acter and text-line-level images. Both datasets are available freely and can
Figure 4.1: Sample synthetically generated Amharic character images.

All these character images are generated using different degrada-
tion levels
be also obtained from the author, for research purposes. Sample character
images from this dataset are illustrated in Figure 4.1. The following section
presents the details of the ADOCR database.
4.3.1 Amharic Character Image Dataset
Despite their numerous attempts to do so, researchers have not been able to
prepare Amharic datasets that are large enough to train deep learning models.
In addition, none of them are made the datasets publicly available. Therefore,
to advance the OCR research for Amharic script, a new dataset of Amharic
characters is proposed.
This dataset is organized into two groups. The first group, Type-I, consists
80,000 Amharic character images with the corresponding Ground-Truth (GT)
texts. The character images and their corresponding GT, in this group, are in
the format of png and txt file respectively. Since it has no explicitly separated
training and test dataset, researchers are free and should split both the training
and test sets accordingly.
However, the second group of the character level dataset contains 72,194
sample Amharic character images as a training and 7800 sample images as
a test set. In this dataset, Type-II, both the character images and the corre-
sponding GT are stored in numpy file format. Moreover, the images are Grey-
level with a size of 32 by 32 pixels and generated synthetically using, OCRo-
pus [Bre08], an open-source OCR framework. Except about 2006 distorted
Amharic character images, which are removed from Type-I, both Type-I and
Figure 4.2: Fidel Gebeta. Basic Amharic characters are arranged in row-
column-wise (33 × 7) order; read from left to right: The first col-
umn gives the consonant characters while the next six columns are
derived by combining the consonant with vowel sounds.
Type-II datasets contain the same character images. In this dataset, we only
considered 231 basic Amharic characters which are arranged in a matrix-like
structure, called Fidel Gebeta, and illustrated in Figure 4.2.
4.3.2 Amharic Text-line Image Dataset
Many research works on Amharic script recognition so far were focused on

character level recognition as a result no attempts have been made by re-
searchers to prepare a text-line database for Amharic OCR; thus the first at-
tempt is made in this work. Amharic text-line image dataset, part of ADOCR
database, contains 337,337 Amharic text-line images composed of the two
commonly used fonts (Visual Ge’ez and Power Ge’ez fonts) of Amharic script.
Figure 4.3: Synthetic Amharic text-line image generation pipelines. The

textual input along with desired font type are goes through the
OCRopus-linegen and then the text-line images and their corre-
sponding Ground-Text GT texts are generated. Variations of each
text-line image are depends of degradation parameters and fonts
considered during data generation.
These text-line images have two parts. The first part is synthetically generated
Amharic text-line images while the second part is a printed Amharic text-line
images.
Following the popularity of using artificial data in computer vision and pat-
tern recognition, the shortage of training datasets is, now, partially solved. A
similar approach is employed, in this thesis, to address the problem of limited
labeled Amharic script dataset. During synthetic data generation, to make the
synthetic data closely similar to the scanned documents that represent real-
world data (such as degradation, font, and skew), we use different parameters
from the OCRopus framework such as threshold with values between 0.3 and
0.5, jitters with the value of 0.5 and so on.
As shown in Figure 4.3, the utf-8-encoded text line and ttf -type font-files
are used as the inputs of OCRopus-linegen engine and generate the text-
line image with corresponding GT texts. Since the OCRopus-lingen engine
doesn’t generate the character-level image directly, we prepare a text file that
Figure 4.4: A sample scanned page of Amharic document. This is a two

column sample image among a collection of printed Amharic doc-
uments.
consists of a character per line first and then feed this text file to the engine
for character-level image generation. Furthermore, there are some necessary
changes regarding the OCRopus framework code so as to adapt for documents
written in Amharic script.
Even though using synthetic data is common in most pattern recognition

and OCR research, it is known that the synthetic data doesn’t represent all the
real-world artifacts. Moreover, the OCR model trained with the synthetic data
lacks generalization of the real data. The usage of real-world data, if available,
is always a choice in the field of OCR. For Amharic OCR, there is no printed
text-line image dataset that could be used to train the OCR model and explore
real-world challenges so as to develop an effective digitization system for
Amharic script. In this regard, we are motivated to develop a compressive
Amharic text-image dataset for benchmarking state-of-the-art methods.
During printed text-line image dataset preparation the contents of the doc-
ument are an important factor; thus the collection of contents should have to
OCRopus-nlbin
OCRopus-gpageseg
Scanned page
segemnted Amharic text-line images
Figure 4.5: A typical diagram of OCRopus page segmentation-engine.

The OCRopus page segmentation engine produces a segmented
text-line image from a scanned page of documents.
represent the real facts in Amharic script. In this regard, an electronic copy
of Amharic documents like letters, newspapers, books, and magazines which
are written with 12 point pixel including cultural, economical, political, and
religious contents was collected. Once the documents are collected and or-
ganized, the printing process is done using a laser jet printer with the usual
setting used in Amharic document printing process.
The next step, after printing, is image acquisition which can be done using
different technologies like cameras, scanners, and other mobile-based tech-
nologies. In most document digitization, the scanner-based acquisition is con-
sidered as a standard approach and abundantly used for scanning books, mag-
azines, and other textual resources. The scanned contents of printed Amharic
texts are acquired by scanning 721 pages of printed Amharic documents using
a flatbed scanner with 300dpi. Sample scanned page of Amharic document is
shown in Figure 4.4.
For page image binarization and text-line segmentation, a binarization and

line segmentation techniques from OCRopus framework are employed. The
overall page binarization and line segmentation procedures are illustrated in
figure 4.5. The errors observed during text-line segmentation are annotated
and aligned with its corresponding GT text manually. This technique can
be used as an alternative solution for printed dataset preparation and it re-
Figure 4.6: Amharic characters and punctuation marks. A representa-

tive of each character and punctuation mark are presented in
the ground-truth of the training and test samples of the ADOCR
database.
Figure 4.7: Text-line image length normalization. An image in (a) has a

size larger than 48 × 128 while image in (b) is less than that.
Therefore, both the images are scaled proportionally. Images at
the left are random text-line images from the dataset while the
right side images are resized to 48 × 128. An image with empty
space at the right side of the image (b) is part of the target image
filled with white color. By doing image size normalization enables
us to train the network with a batch of data shuffled in every epoch.
duces the time and effort required during traditional text-line image annota-
tion which is usually done by transcribing each text-line manually by human
operators or language experts.
From the total set of text-line images, 40,929 text-lines are printed text-
lines written with Power Ge’ez font, 197,484 and 98,924 text-line images are
synthetically generated text-lines with Power Ge’ez and Visual Ge’ez fonts re-
spectively. In the GT texts there are 280 unique Amharic characters and punc-
tuation marks which are depicted in Figure 4.6 and all characters that exist in
the test dataset also exist in the training set. The text-lines in the ground-truth
of the ADOCR dataset have a minimum sequence length of 3 and a maximum
length of 32 characters. This dataset totally consists of 851,711 words that
are composed of 4,916,663 characters, where 269,419 characters are from the
test datasets and 4,647,244 characters are from the training dataset.
Since the text-line images have different lengths and the implementation
deep learning framework, Keras, used in this thesis requires a limited se-
a b c
Figure 4.8: Sample Amharic text-line images from ADOCR database. All
images in this figure are normalized to a size of 48 by 128 pix-
els. (a) Printed text-line images written with the Power Ge’ez font
type. (b) Synthetically generated text-line images with the Visual
Ge’ez font type. (c) Synthetically generated text-line images with
the Power Ge’ez font type.
quence length to optimize GPU parallelization, all text-line images are resized
to a width of 128. During text-line image length normalization, first, we cre-
ate a white target image of size 48 × 128. Then we copy each text-line image
into the white target image, as shown in Figure 4.7. Now all the text-line im-
ages are resized to an equal and fixed length which is easy for batch training
where the batch size is greater than one.
The synthetic text-line images, in this database, are generated and stored at
a different level of degradation. The text-line images and their GT text files
in the training set are in the order of printed, synthetically generated images
with Power Ge’ez and Visual Ge’ez fonts respectively. All images and their
corresponding GT text are in the format of numpy file. Images in a test-set are
organized as a separate numpy file with their corresponding GT texts each.
Sample printed and synthetically generated text-line images taken from the
ADOCR database are shown in Figure 4.8. The test dataset contains a total of
18,631 randomly selected text-line images where 2907 of them are scanned
printed, 9245 and 6479 of them are synthetically generated text-line images
with Power Geez and Visual Geez fonts types respectively.
4.4 Chapter Summary
In this chapter, a baseline Amharic OCR database, called ADOCR, is pre-

sented. This database is composed of printed and synthetically generated im-
ages of Amharic documents. The ADOCR database consists of two types
of datasets; the first dataset is composed of images at character-level while

the second database consists of text-line level images. Besides, the ADOCR
database consists of two common Amharic fonts (Power Ge’ez and Visual
Ge’ez) To develop these datasets two approaches are employed, one is to gen-
erate images synthetically using different degradation levels and font types
while the second approach is to prepare the scanned printed text-line images.
The main contribution of this chapter is the creation of the ADOCR database,
an Amharic OCR database that consists of synthetically generated images of
Amharic characters and text-lines. It also contains scanned printed Amharic
text-line images with their corresponding ground-truth texts. The contribu-
tion of this new Amharic database is not only to the field of Amharic script
recognition but also to many other document image analysis tasks, in general.
In the next chapters, of this thesis, we provide the details of OCR models pro-
posed for Amharic script recognition and experimental results recorded using
this OCR database.
Chapter 5
Recognition of Amharic Characters
Nowadays, it is becoming increasingly important to have digital documents

for easy access to information, efficient data storage, and retrieval. Therefore,
the use and applications of OCR systems have been developed and widely
applied for the digitization of most languages. In this part of the thesis, an
overview of Amharic script and CNN-based Amharic character recognition
systems are provided. This chapter has two sections: the first section of
this chapter states the background of Amharic script and attempts done to
develop a character recognition model for basic Amharic characters based
on a standard CNN architecture. The second section presents a novel CNN-
based Amharic character recognition network architecture which is designed
by leveraging the grapheme of Amharic characters in Fidel-Gebeta and by
taking the concept of multi-task learning.
5.1 Amharic Character Recognition Using

CNNs
In this section, we introduce the first attempt for Amharic character image
recognition based on deep learning benchmarks. A standard Convolutional
Neural Network (CNN) has been employed for the recognition of basic Amharic
characters. The proposed method is trained with synthetically generated Amharic
character images so as to replace the shortage of training data and the perfor-
mance of the model evaluated using both synthetically generated and printed
Amharic characters. The detailed implementation procedure and related works
are provided in the following section.
53
54 Recognition of Amharic Characters
5.1.1 Introduction
Amharic is an old Semitic language that has been and is widely spoken in
Ethiopia. It is widely used in the country and serves as the official work-
ing language of the state. Most historic and literary documents are written
and documented in this language. Despite the importance of the language,
scanned image documentation and retrieval in Amharic script is still a tedious
process. Technologies that can automate the process of changing this scanned
image into a full and editable document will be of enormous use in many sec-
tors of Amharic speaker nation’s economy. However, unlike the well-known
Latin scripts such as English and Fraktur and Non-Latin script such as Ara-
bic scripts, there is no workable and effective off-the-shelf OCR software for
Amharic script recognition. Over the last few decades, a limited amount of
research efforts have been done for the recognition of Amharic scripts, where
almost all these attempts are based on classical machine learning techniques.
Most published research works on Amharic OCR [Tef99, Ass02, MJ07,

Mes08], relied on traditional and statistical machine learning-based techniques
such as Support Vector Machines and Multi-layer Perceptrons. However, sig-
nificant progress has been made in OCR techniques for many other scripts
[KJEY20, BUHAAS13, YBJL12, BCFX14]. Adoption of recent tools and
algorithms, to Amharic scripts, will be greatly advance the state-of-the-art in
Amharic OCR.
OCR has been and is widely used for many scripts as a method of digi-
tizing printed texts which can be electronically edited, searched, stored more
compactly, displayed on-line, also used to facilitate the human-to-machine
and machine-to-machine communication such as machine translation, text-
to-speech, key data, and text mining [YBJL12, Mes08, GDS10]. Previous
research works on Amharic character recognition follows a long procedure
which starts by segmenting input image into lines, characters and then fol-
lowed by feature extraction and selection. The selected features are used as
an input for a classifier which causes recognition lattice [YBJL12, Tef99] and
also requires more time and techniques during preprocessing and feature ex-
traction.
To optimize the efficiency of the OCR model and minimize information

loss during feature extraction, we adopt deep learning-based techniques from
the existing work [AT17] for Amharic character recognition. Convolutional
neural network (CNN), is selected for its less number of steps and com-
putational resources which have been used for statistical feature extraction
1. Amharic Character Recognition Using CNNs 55
[Bre08, YBJL12, ZHZ17]. CNNs take the advantages of GPU-days of com-

putation to train on large-scale datasets and it is also widely applied for char-
acter recognition [ZHZ17], image recognition [SZ14] and numerals recogni-
tion [AT17]. CNNs enable to automatically extract the salient features, which
are invariant for a certain degree to shift and shape distortions of the input
characters [BCFX14, ZHZ17]. In addition to this, the use of shared weight
in convolutional layers, the same filter is used for each input in the layers.
Sharing weight across CNN layers, reduces the number of parameters and
improves recognition performance [BCFX14].
As reported in the literature, a small private database of handwritten and

machine-printed images, for most Latin-based scripts such as English and
non-Latin scripts such as Arabic and Amharic, were widely used to train
classical machine learning algorithms and reported many challenges [MJ07,
Tef99, BCFX14]. These researchers are agreed on the nature of the dataset
and most of them recommended that; the use of large and multi-font character
images that have various levels of degradation, for OCR model training, en-
hances the recognition performance. Therefore, we develop an Amharic OCR
system that is capable to recognize Amharic characters written with multiple
fonts and generated at a different level of degradation.
In this research work, a convolutional neural network-based model is pro-

posed with the goal of training the model using a synthetic Amharic charac-
ter image and then applied it for recognition of both synthetically generated
and printed Amharic characters. This model is also a benchmark model for
character-level Amharic OCR. In this regard, different CNN architectures are
trained rigorously so as to achieve very low error rates compared to other
standard systems and methods.
5.1.2 Related Work
For character recognition, various methods have been proposed and high char-
acter recognition performances are reported for the OCR of various scripts
such as English [BCFX14], Fraktur [BUHAAS13], Arabic [GDS10], De-
vanagari [BCFX14] , Malayalam [AMKS15] and Chinese[HZMJ15]. In the
last two decades, many researchers [Tef99, Ass02, MJ07, Hag15] are also
attempted to address the problems in Amharic script recognition.
Dereje [Tef99] proposed a Binary Morphological filtering algorithm for

OCR of Amharic typewritten texts. He recommended that adopting recogni-
tion algorithms that are not very sensitive to the features of the writing styles
of characters helps to enhance the recognition rate of Amharic OCR. Million
[MJ07], conducted empirical research to investigate and extract the attributes
of Amharic characters written in different fonts and then generalize previously
adopted recognition algorithms. Using different test cases, 49.38%, 26.04%,
and 15.75% recognition accuracy were registered for WashRa, Agafari, and
Visual Geez fonts respectively. In addition, Yaregal [Ass02] noted that font
variation is also a problem in designing an OCR system for printed Ethiopic
documents.
Shatnawi and Abdallah [SA15] tried to model a real distortion in Arabic
script using real handwritten character examples to recognize characters. The
distortion models were employed to synthesize handwritten examples that are
more realistic and achieved 73.4% recognition accuracy. Haifeng [ZHZ17],
proposed a deep learning method based on the convolutional neural networks
to recognize scanned characters from real life with the characteristics of illu-
mination variance, cluttered backgrounds, geometry distortion, and reported
promising results.
In summary, the above studies noted that the writing style of characters that
includes font variation, level of degradation, and size are important factors for
OCR. Preprocessing, feature extraction and classification algorithm selection
is also the major step for constructing an effective recognition model. Even
though, many researchers have been attempted to recognize a small set of
Amharic character images using hand-crafted features which is usually not
robust, to the best of our knowledge, no attempts have been made for Amharic
character recognition using deep learning-based techniques. Therefore, there
is still room for improvement of Amharic character image recognition using
state-of-the-art techniques that can reduce preprocessing steps and computa-
tional resources.
5.1.3 Research Methods
This study is concerned with the development of Amharic character image

recognition model using deep learning techniques. For experimentation, about
80,000 synthetically generated Amharic characters and 6715 printed Amharic
character images are employed. The detail description of the character images
is given in chapter 4 section 4.3.1. For generating the synthetic character im-
ages, we employed an open-source generic OCR framework called OCRopus,
which is developed by Breuil [Bre08]. The printed characters are prepared us-
ing the usual printer setting, in Amharic document, and then scanned using
Kyocera TASKalfa 5501i flatbed scanner at a resolution of 300 DPI. Prepro-
cessing steps including binarization, and character image segmentation has
been done using the same tool, the OCRopus framework. The OCRopus OCR
framework and the detail procedure of dataset generation is described in chap-
ter 4 section 4.3.1 and 4.3.2.
5.1.4 The Proposed CNN Architecture
After doing several experiments considering different Convolutional Neural

Network architecture (CNN), the better results reported, in this section, with
CNN architecture depicted in Figure 5.1 and a pseudo-code for training the
proposed model is shown in Algorithm 1. In this architecture, five convolu-
tional layers, two in each block except the last layer, are used consecutively
with a rectified linear unit (ReLU) as an activation function followed by single
max-pooling. we also use the Same padding scheme so as to determine the
output size and two fully connected layers followed by the soft-max function
for the final layers. A 3 × 3 convolutional filter and a 2 × 2 max-pooling are
applied for performing filtering and sub-sampling operations respectively.
For training, we consider Amharic character images that have an input size
of 32 × 32 pixel. To optimize the proposed network model, training is carried
out by using the stochastic gradient descent method with the mini-batches
size of 150 and momentum value of 0.9. A dropout rate of 0.25 was used to
regularize the network parameters. The output of each convolutional block is
fed to the next network block until the probability of the last layer is calculated
where the activation function is Relu. The stochastic gradient descent method
optimizes and finds the parameters of connected network layers that minimize
the prediction loss. The final output of the model is determined by a Soft-max
function, which tells that the probability of the classes being true is computed
using Equation (5.1).
ezj
f (zj ) = PK , for j = 1, ..., k (5.1)
k=1 ezk
Where z is a vector of the inputs to the output layer and k is the number of
outputs.
Pooling
ሀ
Pooling
ለ
ሃ.
..
Convolutional-layers FC-layers ፖ
Figure 5.1: A CNN-based OCR model for Amharic character image

recognition.
Algorithm 1: Pseudocode for training the proposed model

Input: Model, T, t Model= Keras sequential model, T=training data-set, t=test data-set
Output:trained model
Start: adding convolutional and pooling layers
Model.add(x Conv y, T, relu) x= feature-maps, y= filter size, conv=convolution layer
Model.add(x Conv y, T, relu)x= feature-maps, y= filter size, conv=convolution layer
Model.add(xPy) x= is pooling size, y= stride, p=pooling layer
... here adding the 2 convolutional followed by pooling layers and then Conv
Model.add(xFC) x=number of neurons, FC= fully connected layer
Model.add(NC) NC=number of class
Model.add(softmax)
Model.compile(Optimizer, accuracy measure)
Model.fit(T, epochs, validation-split)
Stop
5.1.5 Experimental Results
The implementation is based on Keras Application Program Interface (API)

with TensorFlow backend. Experiments are run following the network archi-
tecture and classification settings introduced in section 5.1.4, then the training
and validation loss observed during experimentation is depicted in Figure 5.2.
During experimentation, we train our model using scholastic gradient descent
optimizer for different epochs, compared the results, and rerun our model on
more epochs with selected parameters, and the most promising validation ac-
curacy was recorded at 100th epoch using 70%, 10% and 20% of training,
validation and test sets of synthetically generated Amharic character image
dataset respectively.
Figure 5.2: Training versus validation loss of CNN-based OCR model.
Once the recognition model is developed, the performance of the proposed

model is evaluated against the test datasets. The first test dataset contains syn-
thetic Amharic character images generated using OCRopus OCR-framework
with the most common Amharic font called Visual Geez. The second test sets
are printed Amharic character images which are prepared separately so as to
evaluate the performance of the model trained with the synthetic one. This
dataset consists 6,715 printed Amharic characters written with the different
Amharic fonts such as Power Geez, Visual Geez, Agafari and WashRa. Based
on the results recorded during experimentation, 96.02 % of printed charac-
ters are correctly predicted while 92.71% of synthetically generated character
images are correctly predicted.
An investigation of experimental results indicated that most of the charac-

ters are incorrectly recognized due to shape similarity while the others are in-
correctly recognized even with no similarity between them. The main reason
to have lower recognition performance in the synthetically generated dataset
is the nature of the character images (i.e most of the synthetic character im-
ages which are incorrectly recognized are deformed characters). On the other
hand, compared to the synthetic character images the printed character images
have relatively better image quality. In this section, we only consider the 231
basic Amharic character sets.
5.1.6 Recognition Performance Comparison
To the best of our knowledge, there is no publicly available dataset for Amharic
script recognition and all researchers have been done experiments by prepar-
ing their own private datasets. Therefore, it may not be directly relevant to
compare the current Amharic character recognition model with other works
which are performed under different experimental conditions and datasets.
However, the comparison between these works can be used as an indication
of research progress for Amharic character image recognition. The result of
the current Amharic character recognition model and the results of previous
classical machine learning-based research works are shown in Table 5.1.
Table 5.1: Character level recognition accuracy (out of 100%) and perfor-
mance comparison with related research works.
Dataset size Document type Classifier Accuracy

Dereje [Tef99] 5172 Typewritten ANN 61%
Million [MJ07] 7680 Printed SVM 90.37%
Yaregal [Ass02] 1010 Printed ANN 73.18%
Fitum [Dem11] 1500 Printed SVM 91%
Ours 80000 Synthetic CNN 92.71%
The second column (data size) refers to the number of character images used for
experimentation
5.1.7 Conclusion
We experimentally developed an Amharic character image recognition model

using a convolutional neural network and the proposed Amharic character
recognition model outperforms other state-of-the-art methods proposed for
Amharic script recognition. The convolutional neural network tends to work
better with raw input pixels rather than features or parts of an image. During
dataset preparation, we consider only 231 basic Amharic character images. In
our dataset, each character exists 376 times on average. Once we developed
the recognition model with the synthetically generated Amharic character im-
ages, we used both synthetically generated and printed test sets of Amharic
character images. The printed characters are about 10% of the total dataset
and achieved average recognition accuracy of 96.02% while 92.71% of the
synthetic character images are correctly predicted. As we observed from the
2. Amharic Character Grapheme-based OCR 61
empirical results, confusion among visually similar character across rows and
columns were the major challenges. Therefore, we need to develop an alterna-
tive method to handle such confusion among these visually similar Amharic
characters. The next section, Section 5.2, presents an alternative OCR ap-
proach which is designed by leveraging the grapheme of basic Amharic char-
acters on Fidel-Gebeta.
5.2 Amharic Character Grapheme-based OCR
This section presents a novel CNN-based approach for Amharic character

image recognition, which is designed by leveraging the structure of Amharic
grapheme and taking the advantage of multi-task learning. Of 317 Amharic
characters [MJ07], the most widely used characters in Amharic script are ba-
sic characters which we consider, in this section, are arranged in a matrix
structure with a 34 by 7 row-column order and illustrated in Figure 4.2. These
Amharic characters could be decomposed into a consonant and a vowel. As a
result of this consonant-vowel combination structure, Amharic characters lie
within a matrix structure called Fidel-Gebeta. The rows and columns of Fidel-
Gebeta correspond to a character’s consonant and the vowel components, re-
spectively. However, the character ª, on the last row of the Fidel-gebeta that is
a lately introduced Amharic character, is not considered in our dataset. Thus,
we only use the 33 by 7 matrix of Amharic character distribution of Fidel-
Gebeta..
In section 5.1, we noted that the characters shape similarly across rows,
and having large classes affect the performance of the OCR model developed
for Amharic character recognition and results from the OCR being complex.
However, there is a point to be considered so as to overcome class size com-
plexity and character similarity across rows of the Fidel-Gebeta. Based on
the structure of Fidel-Gebeta, dividing the number of large classes into two
smaller subclasses (row-wise and column-wise classes).
To the best of our knowledge, no attempts have been made to take this ad-
vantage and recognize a character based on the row-column order of Amharic
character in Fidel-Gebeta which has motivated us to work on it. Therefore, in
this section, we propose a method called Factored Convolutional Neural Net-
work (FCNN) for Amharic OCR. FCNN has two classifiers that share a com-
mon feature space before they fork-out at their task-specific layer, where the
two task-specific layers are responsible to detect the row and column compo-
nents of a character. The following section first presents a generic character-

level OCR model developed, using classical and deep learning algorithms,
for various scripts including Amharic, and then the proposed script-specific
system architecture designed for Amharic OCR is described in detail.
5.2.1 Character-level OCR
Extensive research efforts have been made to develop an OCR model for vari-
ous scripts and most of these scripts, now, have robust OCR applications while
other low-resource scripts, like Amharic, are still remained relatively unex-
plored [BHS18]. This section presents various character-level OCR works
that have been done, using either classical machine learning or deep learning-
based techniques, for different scripts including Amharic.
The first work was done on Amharic OCR by Worku in 1997 and he adopted
a tree classification scheme built by using topological features of a char-
acter [Ale97]. It was only able to recognize an Amharic character written
with W ashera font and 12 point type. Then other attempts have been made
ranging from typewritten [Tef99], machine-printed [MJ07], Amharic braille
document image recognition [HA17], Amharic document image recognition
and retrieval [Mes08], numeric recognition [RRB18] to handwritten [AB09].
However, all these attempts are based on classical machine learning tech-
niques and trained with a limited number of private datasets. In addition, sev-
eral other works have been made based on CNN-based methods together with
the classical approaches. Maitra et al [MBP15] proposed a CNN as a train-
able feature extractor and SVM classifier for numeric character recognition
in multiple scripts. A similar study, Elleuch et al [ETK16] employed a CNN
as a feature extractor and SVM as recognizer to classify Arabic handwritten
script using the HACDB database.
Even though attempts have done for Amharic OCR so far are based on
traditional machine learning techniques, in section 5.1 of this thesis, a CNN-
based architecture for Amharic character image recognition is presented. Be-
sides, all of these methods mentioned above consider each character as a sin-
gle class [MJ07, Tef99, BHS18] but not consider the row and column order
location of the character in Fidel-Gebeta. The problems of these approaches
are; treated structurally similar characters in a different class and using a large
number of classes which leads the network to be complex. It is also diffi-
cult to obtain a good performance when considering each structurally similar
character as a class and without explicitly using the shared information, in
Fidel-Gebeta, between characters. The following section presents the pro-

posed FCNN model.
5.2.2 Research Methods
There are very few databases used in various works on Amharic OCR reported
in the literature. As reported in [Tef99], the authors considered the most fre-
quently used Amharic character with a small number of samples, which are
about 5172 core Amharic characters with 16 sample images per class. Later
work done by million et al [Mes08] uses 76800 character images with differ-
ent fonts type, sizes with 231 classes. Other researchers’ work on Amharic
OCR [Ale97, AB09, RRB18] reported as they used their own private database,
but none of them made their database publicly available for research purpose.
In this section, we use the same Amharic character image database em-
ployed in section 5.1 that contains 80000 sample Amharic character images.
In this database, there are many deformed and meaningless character images.
Therefore, for our experiment, we removed about 2006 distorted character
images from the database. We also labeled each character image in a row-
column order following the arrangements of Fidel-Gebeta.
Sample Amharic characters from Fidel-Gebeta and shape difference across

their corresponding vowels in row-wise order is depicted in Figure 5.3.
5.2.3 Multi-Task Learning
This section describes the Multi-Task Learning (MTL) paradigm, one of the
central training paradigms in machine learning, that can learn and solve mul-
tiple and related tasks at the same time while exploiting commonalities and
differences across all the tasks [DCBC15, DHK13]. Learning multiple re-
lated tasks simultaneously is inspired by the biological nature of learning in
humans, where we often apply the knowledge we have already acquired by
learning-related tasks from the environment so as to learn new tasks in the
same and/or different environment. MTL has been applied in many domains.
Zhang et al [ZLLT14] propose a convolutional neural network to find facial
landmarks and then recognize face attributes. Kisilev et al [KSBH16] present
a multi-task CNN approach for detection and semantic description of lesions
in diagnostic images. Yang et al [YZX+ 18] introduce a multi-task learning al-
gorithm for cross-domain image captioning. The common part in all these re-
Figure 5.3: Shape changes in row-column structure of Fidel-Gebeta. We

call the first entries of each row the base character. Each row has
its own base character, as a result it has a unique shape. Each
column has a unique ”pattern” of modification it applies to the
base characters. For instance the third column marked by green
box adds a dash at the right bottom, the fourth column marked
by blue box makes the left leg of the character shorter, the fifth
column marked by violet box adds circle at the right leg and so
on.
searchers’ work is to use the same parameters for the bottom layers of the deep
neural network while task-specific parameters at the top layers. As explained
by Rgyriou et al [AEP07], learning multiple related tasks simultaneously has
been often shown significantly better performance relative to learning each
task independently. Compared to training tasks separately, MTL improves
learning efficiency and prediction accuracy for the task-specific models. MTL
provides various services such as implicit data augmentation, attention focus,
and eavesdropping [AM90].
MTL could follow different training strategies with the objective of im-
proving the performance of all the tasks [AEP07, ZY17]. Joint training is one
of the most commonly used training strategies, where the model uses hard pa-
rameter sharing and tried to optimize a single joint loss L which is computed
from each task using,
XN
L= li (5.2)
i=1
where li is the loss of a task Ti and N is the total number of tasks involved
in the training procedure. The character recognition model proposed in this
section is designed by taking the advantages and learning strategies of MTL.
0
1
Row-wise order
2
.
Max-pooling
FC1
(33x1)
Max-pooling
.
.
32
Column-wise order
1
2
Convolutional-layers 3
FC2
(7x1)
.
FC-layers .
.
6
Figure 5.4: Factored CNN architecture:. This method has two classifiers
which share a common feature space at the lower layer and then
fork-out at their last layers. FC1 and FC2, at the last layer, repre-
sent the fully connected layers of row and column detector having
33 and 7 neurons that corresponding to the number of classes for
each task respectively.
The details of the proposed system architecture designed, based on MTL, for
Amharic character image recognition are described in the next section.
5.2.4 The Proposed FCNN Architecture
The overall framework of the proposed approach is shown in Figure 5.4. In

this architecture, we employ six 3×3 convolutional layers with rectified linear
unit (ReLU) activation function and two stacked fully connected layers in
common for both classifiers and fork-out by adding one fully connected layer
with soft-max activation function for each classification branch which has 33
neurons for row detector and 7 neurons for column detector branch. The final
output of each task is determined by a Soft-max function using Equation (5.1).
In the proposed method, we jointly trained the two tasks (row detector and
column detector) by taking the advantage of multi-task learning. In such a
way, when we train the network, the parameters of the row detector layer
should not change no matter how wrong we get in the column detector, and
vice versa, but the parameters of the shared layer changes with both tasks
since we only optimize a single joint loss which is formulated as
ljoint = lr + lc (5.3)
Where lr and lc denotes different losses of row and column detector which
can be computed by Equation (5.4).

C
X
loss = − ti log(si ) (5.4)
i
Where C is the number of class, ti is ground truth for each class and si is a
score of each class calculated by Equation (5.1).
The overall character recognition accuracy of the model is determined by

the recognition accuracy of both tasks. In this case, a single character is cor-
rectly recognized when the row and column components of a character are
correctly detected by both classifiers simultaneously. The index of a charac-
ter in Fidel-Gebeta can be computed from a row and a column component
values of a character as:
Ic = a × 7 + b (5.5)
where Ic is the character index, a and b are the row and column order values
of the character in Fidel-Gebeta.
The total number of characters correctly recognized by the proposed method

depends on the number of times in which both row and column detector net-
works are correct, for each character, simultaneously. Then the recognition
performance of the proposed model is measured using accuracy and can be
formulated as Equation (5.6).
T
n(Sr Sc )
A= × 100 (5.6)
t
T
where A is overall accuracy, is set intersection, Sr and Sc are samples
correctly recognized by the row and column detector respectively, t is number
T
of test samples and n(Sr Sc ) is total number of samples that are mutual
detected as correct by the row and column detector.
Experiments were run following the network architecture introduced in sec-

tion 5.2.4. The input sample character is recognized in terms of its row and
column components, as shown in Figure 5.5.
We carry out our experiments on the Amharic database employed in section

5.1 by removing distorted and meaningless images. Then 77,994 character
images that contain on average 2364 samples from each row of Fidel-Gebeta
Figure 5.5: A typical diagram that shows the prediction of FCNN model.
For an input character image, the row component detector rec-
ognizes zero and the column component detector recognizes three
then the model output becomes the character (ƒ) which is located
at row zero and column three of the Fidel-Gebeta
were used for training and testing. The proposed FCNN architecture contains
two multi-class classification problems (the row and column detector) and the
convolutional layers have 32 filters of size 3 × 3. Every two blocks of convo-
lutional layers are followed by a 2 × 2 max-pooling. The two fully connected
layers have 512 neurons each while the last forked layers have a size corre-
sponding to the number of unique classes in each specific task, in our case
33 for row detector and 7 for column/vowel component detector. Considering
the size of the dataset, we train our network using the different batch sizes and
the best test result is recorded with the batch size of 256 and Adam optimizer
running for 15 epochs. Based on the results recorded during experimentation,
Table 5.2: Performance of prior works (CNN) and proposed method(FCNN)
Dataset size Recognition Accuracy

CNN [BHS18] 80,000 character 92.71%
FCNN 77,994 Row-column 94.97%
96.61 % of the row and 95.96% of the column components of the characters
are correctly detected by the row detector and column detector respectively.
An overall character recognition accuracy of 94.97% was recorded and the
training converges faster since we reduce the number of classes from 231 to
40 classes (33 row-wise and 7 column-wise classes). As can be observed in
Table 5.2, we achieved better recognition performance compared with works
done on Amharic OCR using 231 classes. The value ’character’ in the third
column of Table 5.2 refers to the recognition made with 231 classes while
’row-column refers to recognition with two smaller subclasses, which is done

by dividing 231 classes into (33 row-wise and 7 column-wise classes).
5.2.6 Conclusions
In this section, we have introduced a novel CNN-based method, for Amharic

character image recognition. The architecture consists of two classifiers: the
first is responsible to detect a row component and then the second is responsi-
ble to detect the column component of a character in Fidel-Gebeta. A charac-
ter is correctly recognized when both detectors are correctly identified in its
row and column components. We train and evaluate our Factored CNN-based
recognition model on a synthetically generated Amharic character database
that has 77994 sample Amharic character images. Empirical results show
that our proposed method achieved state-of-the-art performance.
Furthermore, our proposed method minimizes the problems of class size
complexity and row-wise character similarly by considering the structural
arrangement of Amharic characters in Fidel-Gebeta and then splitting the
classes into two smaller subclasses. As observed from the empirical results,
the confusion between visually similar characters such as X and ª , † and
˜ which are found in row 1 and 6, 7 and 33 of Fidel-Gebeta respectively is
greatly reduced and therefore row-wise character similarity is not now a chal-
lenge for this method. Besides, the performance of the new model is far better
than the standard CNN-based OCR model.
5.3 Chapter Summary
This chapter discusses the application of CNN-based OCR methodology for

Amharic character image recognition. In this regard, this chapter introduces
a novel methodology for Amharic character recognition. A Factored CNN-
based character recognizer is designed by following the architecture of multi-
task learning and has been trained to learn class associations of individual
characters based on their grapheme on Fidel-Gebeta. Experimental results
show that the new factored CNN-based model outperforms the other CNN-
based model for Amharic character image recognition. It also minimizes the
problems of class size complexity and row-wise character similarly that have
been limiting the recognition performance of the OCR models designed for
Amharic character recognition. In general, empirical results on both synthetic
and printed Amharic characters, convolutional neural network-based OCR

models are significantly better than classical machine learning approaches.
In addition to these character-level Amharic OCR methods, the next chapters
of this thesis describe various state-of-the-art sequence-to-sequence methods
proposed and implemented for Amharic script where the recognition is at a
text-line level.
Chapter 6
Amharic Script Recognition using
CTC Sequence Modeling
In the classical OCR approach, different handcrafted features have been ex-
tracted from segmented components of a given script, and then pattern clas-
sification algorithms were employed for recognition. However, contempo-
rary deep learning techniques can learn the distinguishing features automat-
ically; thereby eliminating the manual feature extraction steps. Sequence
learning is based on the sequence-to-sequence matching principle where a
neural network is trained such that both the input-output pairs are comprised
of variable length sequences. This part of the thesis overviews state-of-the-art
segmentation-free OCR techniques and contains two sections. Each section
provides a detailed analysis of the corresponding OCR techniques and reports
their recognition performance on Amharic text-image recognition.
Further this chapter is organized as follows: Section 6.1 describes the use
of LSTM-CTC-based networks for Amharic OCR. Section 6.2 presents the
details of a hybrid model based on the CNN and LSTM networks for Amharic
text-line image recognition. Section 6.3 concludes the chapter with a short
summary.
6.1 Amharic Text-image Recognition with

LSTM-CTC Networks
This section provides necessary details about the LSTM-based OCR method-
ology that has been used for Amharic text-image recognition and detail anal-
71
72 Amharic Script Recognition using CTC Sequence Modeling
ysis is also presented based on the results obtained from experimentation.
6.1.1 Related Work
Over decades of research in OCR, various techniques have been proposed to

carry out the text-line image recognition task for multiple scripts and even
most of the scripts, now, have workable OCR systems. Some of these works
are presented in chapter 5 of this thesis. Further, deep learning-based sys-
tems have demonstrated ground-breaking performance on image recognition
[HZRS16], sequence prediction and labeling [LGF+ 07, GLB+ 08]. Based
on deep neural network, many segmentation free OCR techniques have been
studied and impressive progress has been made for many Latin and Non-Latin
scripts of both printed and handwritten such as Bidirectional Recurrent Neural
Network with Long Short Memory (LSTM) architecture for online handwrit-
ten recognition [LGF+ 07], Convolutional Recurrent Neural Network (CRNN)
for Japanese handwritten recognition [LNNN17], segmentation free Chinese
handwritten text recognition [ML15], a hybrid Convolutional-LSTM for text
image recognition [Bre17], Multidimensional-LSTM for Chinese handwrit-
ten recognition [WYCL17], combined Connectionist Temporal Classification
(CTC) with Bidirectional LSTM (BLSTM) for unconstrained on-line hand-
written recognition [GLB+ 08].
Research attempts, on Amharic OCR, have been made to develop Amharic
OCR starting from 1997 using different statistical machine learning tech-
niques [MJ07] [Ale97, CH03, Tef99]. Following the success of deep learning,
other attempts have also been made for Amharic character recognition em-
ploying Convolutional Neural Network (CNN) [BHS18, RRB18, BHL+ 19b,
GSTBJ19]. As presented in chapter 5 in detail, nearly all attempts have per-
formed segmentation of text image into character level which directly affects
the performance of the OCR. The only exception is Assabie [AB11], pro-
posed an HMM-based model for offline handwritten Amharic word recogni-
tion without character segmentation using structural features of characters as a
building block of the recognition system. Further, the research attempts done
for Amharic OCR so far neither use enough training data nor considering all
possible characters of Amharic script in their experiment [BHS18].
There are many effective off-the-shelf commercial and open source OCR
applications for many languages including the functionality of a ground truth
generator from the exiting printed text and even a synthetic handwritten train-
ing data generator OCR engines [GSR15, RBO17]. The effectiveness of var-
1. Amharic Text-image Recognition with LSTM-CTC Networks 73
ious open-source OCR engines was evaluated on 19th century Fraktur scripts
and the evaluation shows as the open-source OCR engine can outperform
the commercial OCR application [RSWP18]. However, the OCR system for
many scripts especially those which are indigenous to the African continent
(such as Amharic) remains under-researched and none of the researchers take
the advantage of deep learning techniques such as LSTM, used for many lan-
guages, for developing Amharic OCR.
Therefore, in this section of the thesis, we propose a recurrent neural net-

work based on Bidirectional-LSTM architecture with Connectionist Tempo-
ral Classification (BiLSTM-CTC) for Amharic text-line image recognition.
In addition, we introduce a technique for the preparation of printed Amharic
datasets.
6.1.2 Database
As explained in chapter 4, there are few datasets of Amharic script reported in

the literature. for example, as reported in [Tef99], the authors considered 5172
images of the most frequently used Amharic characters. A later work by Mil-
lion et al. [MJ07] uses 76,800 character images with different font types and
sizes which belong to 231 classes. Other researchers’ work on Amharic OCR
[Ass02, RRB18, Ale97] reported that they used their own private databases,
but none of them were made their dataset publicly available. Therefore, the
shortage of datasets has been continued as the main challenge and one of the
limiting factors in developing reliable OCR systems for Amharic script to
date. Synthetic data generation has become a surrogate technique and solves
the shortage of training datasets to some extent. However, to date, getting
a real training dataset is one of the limiting factors in developing a reliable
system in the area of image processing and pattern recognition.
Therefore, in this section, Amharic text-line image database presented in

chapter 4 section 4.3.2 is employed for both training and model performance
evaluation. The database consists 337,337 Amharic text-line images. The
detail statistics of this database is summarized in Table 6.1.
6.1.3 The Proposed Model Architecture
The overall network architecture of the proposed model, which is adapted

from the existing work for other scripts [SUHP+ 15, UHAR+ 13], is depicted
Table 6.1: The detail of Amharic text-line images database
Printed Synthetic
Font type Power Geez Power Geez Visual Geez
Number of samples 40,929 197,484 98,924
No. of test-samples 2,907 9,245 6,479
No. of train-samples 38,022 188,239 92,445
No. of unique chars. 280 261 210
in Figure 6.1. In a general pipeline of text-line image recognition, the input

of the model is an image containing a line of texts and the output is the cor-
responding strings of texts. In this architecture, a grey-scale text-line image
used as an input of BLSTM network, and the output of BLSTM is fed into a
soft-max layer which has n+1 nodes, where each node is corresponding to a
label or each unique character in the GT including one blank character. In the
ADOCR database, there are 280 unique Amharic characters. Therefore, the
soft-max outputs 281 probabilities at each time step.
The CTC objective function can automatically learn the alignment between
the input and output sequence [GFGS06], where the CTC loss function is used
to train the network. During training, CTC requires only an input sequence
and the corresponding transcription. The detail formulation of the CTC cost
function explained in chapter 2, section 2.4.1.
Once the probability of label sequence y from an input sequence x is ob-

tained with the CTC forward-backward algorithm proposed by [GFGS06].
We employ the best path decoding method to find a character (C) that has
the highest score from outputs (i) at every time-step and the final recognized
string text can be generated like the example given in chapter 2, section 2.4.1,
without segmentation of the input sequence and formulated as Equation (6.1).
C max = B(arg max(yit )), for t = 1, 2, ..., T (6.1)

i
In all experiments described below, we use the same network architecture and
the performance of the proposed model is calculated in terms of character er-
ror rate which is computed by counting the characters inserted, substituted,
and deleted in each sequence and then divided by the total number of charac-
ters in the ground-truth (see Equation (6.2)).
Figure 6.1: The proposed network architecture. This network consists of

two components. The first layer is Recurrent layers as a sequence
learner and then the soft-max function that predicts a label dis-
tribution of each sequence. The second layer is CTC layer which
takes the soft-max prediction as input and transcribes these pre-
dictions to final label sequences. All the input text-line images
length is normalized to the fixed height of 48 pixels. A raw pixel
value of the 1D sequence is extracted by traversing a 48 × 1 time
window over a text-line image and fed to BLSTM network as input.
A character with the highest score at each time step is computed
using Equation (6.1) and the final string of a text is generated
using the reduction process.
6.1.4 Experiments
Experiments are conducted using, ADOCR, a database of Amharic text im-

ages which consists of both printed and synthetically generated Amharic text-
line images. To select suitable network parameters, different values of these
parameters were considered and tuned during experimentation and the re-
sults reported, in this section, were conducted using Adam optimizer, and the
model is trained to minimize the CTC-based loss. We employed the BLSTM
network with two networks hidden layers size of 128 each and a learning rate
of 0.001. Once we implement the model with Keras Application Program In-
terface (API) on a TensorFlow backend, two experiments are undertaken so as
to evaluate the performance of the model against two types of test datasets of
Figure 6.2: Sample text-line image for illustration of Amharic script. A

word marked by violet box is composed from four individual
Amharic characters.
the ADOCR. The first experiment is testing our model with synthetic Amharic
text-line images generated with Power Geez and Visual Geez fonts. The sec-
ond experiment is used to test how the model, that trained with a synthetic
and part of printed text-line images, works on printed text-line images writ-
ten with Power Geez font type. Sample Amharic text-line image taken from
ADOCR database is illustrated in Figure 6.2.
We randomly split the training and test-set. We use a total of 318,706 train-
ing and 18,631 samples for testing. For validation, we use 7% of randomly se-
lected images from the training dataset. The detail information of the database
used for our experiment is shown in Table 6.1. The training and validation loss
of the proposed model recorded for 50 epochs is depicted in Figure 6.3. The
network was trained for 50 epochs with a batch size of 200 and then a charac-
ter error rate of 4.24% on synthetically generated text-line images with Visual
Geez font, 8.54% on printed text-lines with Power Geez font, and 2.28% on
synthetic text-line images generated with Power Geez font are recorded. The
results recorded during experimentation are summarized in Table 6.2. The
performance of the proposed model is measured using Character Error Rate
(CER) formulated as follows:
!
1 X
CER(P, T ) = D(n, m) × 100 (6.2)
q n∈P,m∈T
where q is the total number of target labels, P and T are the predicted and
ground-truth labels, and D(n, m) is the edit distance between sequence n and
m. Edit distance is the difference between the predicted and ground-truth texts
Figure 6.3: Training versus validation loss of an LSTM-CTC model.
which can be occurred due to the deletion, substitution and insertion of char-
acters during recognition. The results show that using a synthetic dataset is the
Table 6.2: A summary of experimental results (CER) of LSTM-CTC model
Font type CER(%)

Printed Power Geez 8.54
Power Geez 4.24
Synthetic
Visual Geez 2.28
best alternative solution to overcome the shortage of labeled data in general

and specifically for Amharic OCR application which has no publicly avail-
able dataset for research purposes. In addition, the model requires enough
data generated from multiple fonts so as to perform better on varieties of test
data.
Figure 6.4: Sample miss-recognized characters in each text-line image of

the test dataset. (a) Printed text-line images. (b) Synthetically
generated text-line images. The characters marked by redbox are
sample miss-recognized (substituted with other characters), char-
acters marked with green box are sample deleted characters and
characters marked with violet box are sample inserted charac-
ters which are generally called the substitution, deletion and/or
insertion errors, in this thesis, respectively. As we observed dur-
ing experimentation, most of the substitution errors are occurred
on the synthetic test images while most of insertion and deletion
errors are observed on printed test images.
6.1.5 Analysis of the Result
Sample Amharic text-line images and characters miss-recognized during test-

ing are shown in Figure 6.4. In this figure, sample output texts containing
different types of character errors that may occur due to misspelled charac-
ters, spurious symbols, or lost characters are presented. The proposed model
works better on a synthetic test dataset compared to the character error rate
observed on printed test data. The generalization of this model, especially on
printed data, is unsatisfactory. This happens mainly because of a significantly
smaller number of printed text-line images and their nature in the training
samples.
For example, as illustrated on Figure 6.4a, on the top part of the figure, the
2. Deep CNN-RNN Hybrid Networks for Amharic OCR 79
character (Ú ) marked by the green rectangle in the input image is substituted

by the character (g) in the predicted text. On the other hand, the two charac-
ters (y and ¤) marked with the red rectangle are deleted characters. Other
character errors are depicted on the top part of Figure 6.4b; characters (” and
Ý) from the input image, marked with the red rectangles, are deleted charac-
ters, while characters (â and s) in the predicted text, written with a blue
color, are inserted characters.
To the best of our knowledge, there is no similar work which is directly rele-
vant to compare the current work with other Amharic OCR model since other
works are performed under different experimental conditions and datasets.
Therefore, we don’t compare our results with others’ work. In general, the
performance of the proposed method obtained promising results compared to
attempts done on Amharic OCR using segmentation-based traditional meth-
ods and this work will be used as a benchmark for future works on Amharic
OCR.
6.1.6 Conclusions
In this section, we present a recurrent network based on a Bidirectional LSTM

network architecture with a CTC objective function for Amharic text-line im-
age recognition motivated by its success on the task of sequence labeling.
We evaluate our model both on synthetically generated and printed Amharic
text-line images, from the ADOCR Amharic database, and achieved state-
of-the-art results. The next section presents an extension of this model by
integrating CNNs before the LSTMs as a feature extractor.
6.2 Deep CNN-RNN Hybrid Networks for

Amharic OCR
In this section, we introduce an end-to-end Amharic text-line image recog-

nition approach based on hybrid neural network architecture. Motivated by
the recent success of end-to-end learning in pattern recognition, we propose a
model which integrates a feature extractor, sequence learner, and transcriber
in a unified module and then is trained in an end-to-end fashion. In the pro-
posed CNN-RNN hybrid network; the CNN layer is employed at the higher
level as a feature extractor, LSTM networks have been used as a sequence
learner of text-line images where the target labels are the character of Amharic
scripts. Since text-line image recognition is a Spatio-temporal pattern recog-

nition problem, the proposed model, in this section, consists of CNN and
LSTM networks together with CTC. This network architecture is proposed as
an extension work of the section 6.1 with the following summarized contribu-
tions:
• The proposed CNN-based feature extractor module is stacked before

LSTM layer and used to extract automatic features from text-line im-
ages
• We adopt an end-to-end trainable neural network for Amharic text-line

image recognition that achieves state-of-the-art results.
• Based on the experimental results obtained, a detailed analysis of the

dataset is presented.
Our hypothesis in this section is the hybrid of CNN-LSTM networks is better

than LSTM in handling sequential tasks that are both spatially and temporally
deep, like text-line images. This hypothesis is based on the results reported in
the literature; thus CNNs are good at reducing frequency variations in spatial
structure, like images, while LSTM networks are designed for interpreting the
features across time steps and are good at temporal modeling. To justify our
hypothesis, a hybrid CNN-LSTM network is trained by stacking CNN layers
on the front end and then followed by LSTM layers with a fully connected
layer and CTC layer on the output. The proposed model has the flexibility to
be applied to a variety of tasks that involve data with sequential inputs and
outputs. Therefore, in this work, we take advantage of the complementarity
of CNNs and LSTMs by combining them into one unified architecture.
6.2.1 Database
Since there is no other publicly available dataset for Amharic script recog-
nition, to train and evaluate the performance of our OCR model, we use the
same database, employed in section 6.1. In the original dataset, ADOCR, all
text-line images are Greyscale and the sizes of each text-line were 48 by 128
pixels. Considering similar works done in the area, and to reduce computa-
tional costs during training, we resized the images into sizes of 32 by 128
pixels. Sample text-line images that are normalized to a size of 32 by 128
pixel and used for training and testing the model proposed in this section are
illustrated in Figure 6.5.
a b c
Figure 6.5: Sample Amharic text-line images from ADOCR database. All
images in this figure are normalized to a size of 32 by 128 pix-
els: (a) Printed text-line images written with the Power Ge’ez font
type. (b) Synthetically generated text-line images with the Visual
Ge’ez font type. (c) Synthetically generated text-line images with
the Power Ge’ez font type.
6.2.2 The Proposed Hybrid CNN/LSTM Network

Architecture and Parameters
Since CNNs and RNNs are the most suited for image-based problems [HKR15]
and sequence recognition [GVK19] respectively, we proposed a hybrid net-
work framework which is illustrated in Figure 6.6. In this framework, we
employed three modules; the feature extractor, the sequence learner, and the
transcriber module. All three of these modules are integrated into a single
framework and trained in an end-to-end fashion.
The feature extractor module consists of seven convolutional layers which

have a kernel size of 3 × 3, except the one that is stacked on top with a 2 × 2
kernel size; each uses rectifier linear unit (ReLU) activation, four max-pooling
layers with pool sizes of 2 for the first pooling layer, and 2 × 1 for the remain-
ing pooling layers. Strides are fixed to one, and the ‘same’ padding is used in
all convolutional layers. The number of feature maps is gradually increased
from 64 to 512, while the image size is quickly reduced by a spatial pool-
ing to further increase the depth of the network. For an input size N , kernel
K, padding P , and stride S, the output size M at each convolutional layer
can be computed as M =( N −K+2∗P S
)+1. In addition, to normalize the out-
put of the previous activation layer and accelerate the training process, batch
normalization is used after the fifth and sixth convolutional layers.
We also used a reshape function to make the output of the convolutional

layer compatible with the LSTM layer. The sequence learner module is the
f*T
CNN-layers
Reshape
Input-Image LSTM LSTM LSTM
..
.
LSTM LSTM LSTM
..
.
h1 h2 hT
f*1 f*1 f*1
FC+SoftMax FC+SoftMax FC+SoftMax
n*1 n*1 n*1
CTC error calculation and decoding
Output-text
Figure 6.6: The proposed CNN-LSTM model. This network consists of

three components. The first component is the convolutional layer
which acts as the feature extractor module, the second component
is the recurrent network layer which takes the outputs of CNNs
after it is reshaped tof × t and it outputs f × 1 feature vector
for each time step. Then the fully connected networks (FC) with
soft-max function are applied to predict a label distribution, in-
cluding a blank token, on n + 1 number of characters at each
time-step. The third component is a Connectionist Temporal Clas-
sification (CTC) layer which takes the soft-max prediction at each
time-step as input and transcribes these predictions into final la-
bel sequences. All input text-line images length is normalized to
a fixed size of 32 by 128 pixels, and labels are padded with zero
until they reach the maximum sequence length of 32.
Table 6.3: Convolutional network layers of the proposed model and their cor-
responding parameter values for an input image size 32 × 128 × 1.
Network Layers Kernel Size Stride Feature Maps

Convolution 3×3 1×1 64
Max-Pooling 2×2 2×2 -
BatchNormalization - - -
BatchNormalization - - -
middle layer of our framework which predicts the sequential output per time-
step. This module consists of two bidirectional LSTM layers, with the soft-
max function on top, each of which has 128 hidden layer sizes and a dropout
rate of 0.25. The sequential output of each time-step from the LSTM layers
is fed into a softmax layer to get a probability distribution over the n + 1
possible characters. Finally, transcription of the equivalent characters is done
using the CTC layer. The details of the network parameters and configuration
of the proposed model are depicted in Table 6.3 (configuration of CNN layers)
and Table 6.4 (configuration of LSTM layers).
During training, the input text-line image passes through the convolutional
layers, in which several filters extract features from the input images. After
passing some convolutional layers in sequence, we apply to reshape operation
on the output and obtained the sequence of 63 vectors of 512 elements. Then
we feed these 63 vectors to the LSTM network and get its output which also
the vectors of 512 elements. The output of the LSTM is fed into a fully
connected network with the soft-max function which has n + 1 nodes with
a vector of 281 elements. This vector contains the probability distribution of
observing character symbols at each time step. Each symbol corresponds to
a label or each unique character in the ground truth and one blank character
which is used to take care of the continuous occurrence of the same characters.
We employed a checkpoint strategy that can save the model weight to the same
file each time an improvement is observed in validation loss.
Table 6.4: The recurrent network layers of the proposed model with their
corresponding parameter values. The input size of the Long-
Short-Term Memory (LSTM) is a squeezed output of the convolu-
tional layers, which is depicted in Table 6.3.
Network Layers (Type) Hidden Layer Size

BLSTM 128
BLSTM 128
Soft-Max No. class = 281
Once the probability of label sequence y from an input sequence x is ob-

tained with the CTC forward-backward algorithm proposed by [GFGS06], we
employ the best path decoding method, fast and simple, to find a character (C)
that has the highest score from outputs (i) at every time-step; the final recog-
nized string text can be generated using B (see the formulation in chapter 2,
section 2.3 Equation 2.7) without segmentation of the input sequence.
In all results reported in this section, the performance of the proposed
model is described in terms of Character Error Rate (CER) (see Equation
6.2), which is computed by counting the number of characters inserted, sub-
stituted, and deleted in each sequence and then dividing by the total number
of characters in the ground truth.
During training, to select suitable network parameters, different values of

these parameters were considered and tuned, and the results reported in this
section were obtained using an Adam optimizer employing a convolutional
neural network with a feature map that started from 64 and increased to 512,
the BLSTM network with two network hidden layers with sizes of 128 each,
and a learning rate of 0.001.
Based on the nature of the dataset, we conducted three experiments to eval-
uate the performance of the proposed model. Once we trained our network
with the synthetic and some of the printed text-line images, the performance
of the model was evaluated with three different test datasets from the ADOCR
database. In the first and the second experiments, the model was evaluated
with synthetic Amharic text-line images that are generated with the Power
Geez and Visual Geez fonts, respectively. The third experiment was conducted
to evaluate the recognition performance of the OCR model using a printed test
dataset written with the Power Geez font type. For validation, we used 7% of
the training dataset, randomly selected, as proposed in section 6.1. The net-
work was trained for 10 epochs with a batch size of 200. During the testing of
Table 6.5: Experimental results of CNN-LSTM OCR model in CER (%).
Image Type Font Type Test Data Size CER (%)

Printed Power Geez 2907 1.59
Visual Geez 6479 1.05
Synthetic
Power Geez 9245 3.73
the proposed model, character error rates of 1.05% and 3.73% were recorded
on the two tests of the ADOCR database which were generated synthetically
using the Visual Geez and Power Geez fonts, respectively. The model was
also tested with the third test dataset, which consists of printed text-line im-
ages that were written with the Power Geez fonts, and a character error rate of
1.59% was obtained. All empirical results recorded during experimentation
are summarized in Table 6.5.
6.2.4 Discussion and Analysis of Results
This section provides a detailed analysis and description of the dataset first,
and then the results obtained during experimentation are also presented. We
performed repetitive evaluations of our method using the ADOCR benchmark
dataset (see chapter 4, Section 4.3.2), and we also tried to compare with the
state-of-the-art methods on both printed and synthetically generated datasets.
Some of the printed text-line images are not properly aligned with the ground
truth due to the occurrence of extra blank spaces between words and/or the
merging together of more words during printing. In addition, with the syn-
thetic text-line images, a character at the beginning and/or at the end of the
word is missed, which results in misalignment with the ground-truth. To im-
prove the recognition performance, it is important to annotate the data man-
ually or to use better data annotation tools. Of several factors, samples with
wrongly annotated Amharic text-line images and characters are the major fac-
tors causing recognition errors. In general, the recognition errors may occur
due to misspelled characters, spurious symbols, or lost characters, and sample
annotation errors per text-line image are illustrated in Figure 6.7.
The proposed model works better on the synthetic test datasets generated
with the Visual Geez font type, compared to the character error rate observed
on the Power Geez font type and the printed test data. The generalization of
this model, especially on the synthetic dataset generated with the Power Geez
a b
Figure 6.7: Sample wrongly annotated images and GT from the test
dataset. (a) Synthetic text-line image; the word marked with
yellowrectangle is a sample mislabeled word where the first
character (µ) in the GT, marked with a redcircle, is missed in
the input image but it exists in GT. (b) Printed text-line image;
a punctuation mark called a full stop/period, bounded with a
purplerectangle in the input image, is incorrectly labeled as two
other punctuation marks called word separators, indicated by the
green and red text colors in the GT.
font type, is not as good as the Visual Geez one. This happens mainly because
of the significantly larger number of text-line images in the test set and the
nature of the training samples (i.e., the text-line images and the ground truth
are not properly annotated due to the existence of deformed characters and
missing characters in the beginning and/or end of the text-line images during
data generation but not in the ground truth). In addition, text-line images
generated with the Power Geez font are relatively blurred, resulting in poor
recognition accuracy.
6.2.5 Recognition Performance Comparison
As depicted in Figure 6.8 and Table 6.6, the performance of the proposed
model is improved. Compared to an LSTM-based Amharic OCR, the pro-
posed model achieved better recognition performance with a smaller number
of epochs. However, the proposed model took a long time for training. This is
due to the nature of end-to-end learning approaches [Gla17] that incorporate
multiple and diverse network layers in a single unified framework. There-
fore, the proposed model can be assessed and training time may be further
improved by following some other concepts like decomposition [SSSS17] or
the divide-and-conquer approach.
The comparisons among the proposed approach and others’ attempts done
on the ADOCR Database [BHL+ 19a] are presented in Table 6.6, and the per-
formance of the proposed model shows better recognition results on all the
a b
Figure 6.8: Learning loss comparison. (a) The training and validation losses
of LSTM-CTC model recorded for 50 epochs. (b) The training and
validation loss of the proposed model recorded for 10 epochs.
three test sets of the ADOCR database.
Table 6.6: Comparison of test results (CER).
#Text-Lines Image Type Font Type CER (%)

LSTM+CTC [ALT18] * 12 pages printed - 2.12%
LSTM+CTC[BHL+ 19a] 2,907 Printed Power Geez 8.54%
LSTM+CTC [BHL+ 19a] 9,245 Synthetic Power Geez 4.24%
LSTM+CTC [BHL+ 19a] 6,479 Synthetic Visual Geez 2.28%
CNN+LSTM+CTC 2,907 Printed Power Geez 1.56%
CNN+LSTM+CTC 9,245 Synthetic Power Geez 3.73%
CNN+LSTM+CTC 6,479 Synthetic Visual Geez 1.05%
* Denotes methods tested on different datasets.
6.2.6 Conclusions
In this section, we proposed a hybrid CNN-LSTM model as an extended ver-

sion of the OCR model presented in section 6.1 of this chapter, where the
new OCR model consists CNNs as the feature extractor, LSTMs as sequence
learner, and CTC networks as the transcriber in a unified framework. All
these three modules are trained in an end-to-end fashion. The recognition
performance of the proposed model is evaluated using three different types
of test datasets from the ADOCR Amharic database, and it outperforms the
LSTM-CTC based OCR model by a large margin.
6.3 Chapter Summary
In this chapter, state-of-the-art techniques based on LSTM networks have in-

vestigated the newly introduced Amharic text-line images database, ADOCR.
The contribution of this chapter is presented in two sections. First, a bidirectional-
LSTM model is trained with CTC. Second, CNN layers are integrated before
the LSTM layer and the networks are trained as a unified framework in an end-
to-end fashion using the CTC objective function. The performance of both
models is evaluated against the three test datasets from the ADOCR database
and the results show that the performance of the recognition model is much
better in the hybrid of CNN-LSTM network than the usual Bidirectional-
LSTM networks. The generalization of these models proposed, in this chap-
ter, for printed text-images is not as good as the synthetic one. This happens
mainly because of the significantly smaller number of printed text-line im-
ages in the training set. We have found that the majority of the recognition
errors on synthetically generated text-images are due to poor annotation of the
dataset which usually could be deformed or missed characters during text-line
image generation. Further, the CTC loss can be efficiently calculated by the
forward-backward algorithm, but CTC still predicts target characters for every
frame with the assumptions of conditional independence in target characters.
Therefore, the next chapter proposes alternative methods, based on attention
mechanism, for better recognition of Amharic text-line images.
Chapter 7
Attention-based Sequence Learning
for Amharic OCR
RNNs trained with CTC objective function have been employed for text-
image recognition and presented in the previous chapter in detail. It is also
a state-of-the-art approach and widely applied for sequence learning tasks to
date. Recently, another sequence learning technique has been emerged, from
the field of NMT, and has often been shown to improve the performance over
the existing approaches. Therefore, this chapter presents the other technical
contributions of this thesis. These contributions are organized in two subsec-
tions and presented as follows. Section 7.1 describes the Attention mecha-
nism and its implementation detail for Amharic text-image recognition; while
section 7.2 shows the advantage of blending attention mechanism in to CTC
objective function for sequence transcription in general and in particular for
Amharic text-image recognition. Finally, it provides summarized findings
and results achieved from the two text-line image recognition approaches pro-
posed for Amharic script.
7.1 Attention-based Encoder-Decoder Model

for Amharic OCR
Text-image recognition has recently been widely treated as a sequence to se-

quence learning task while traditional segmentation-based character recog-
nition has been applied for a longer time. Later an LSTM-CTC based tech-
niques have been used for the recognition of multiple scripts including Amharic
89
90 Attention-based Sequence Learning for Amharic OCR
script which imparts the valuable spatial and structural information of text-
line images. In this section, we aim to push the limits of such techniques
using Attention-based text-line image recognition which is totally based on
neural networks. Unlike the CTC, Attention-based models explicitly use the
history of the target sequence without any conditional independence assump-
tions. The following section gives a detailed overview and training procedures
of the proposed Attention-based encoder-decoder OCR model.
7.1.1 Introduction
The standard OCR tasks have been investigated based on CNNs and RNNs
[BHL+ 19a] by utilizing the CTC [GLB+ 08] objective function. However,
CTC-based architectures are subject to inherent limitations such as condi-
tional independence assumption, strict monotonic input-output alignments,
and an output sequence length that is bound by the subsampled input length
while the Attention-based sequence-to-sequence model is more flexible, suit
the temporal nature of the text and are able to focus on the most relevant
features of the input by incorporating Attention mechanisms [BCB14].
In addition, Attention enables the networks to potentially model language

structures, rather than simply mapping an input to an output [CCB15]. More-
over, its success in the area of NLP tasks such as neural machine translation
[BCB14, LPM15] and speech recognition [DLY+ 19, WHK+ 17] motivates us
to investigate such models in the context of Amharic text-image recognition.
Unlike the previous sections that use the LSTM-CTC based networks, the
method proposed in this section uses the concept of Attention mechanism
with the following major contributions:
• We propose an Attention-based OCR framework for Amharic text-image

recognition for the first time.
• The proposed method is trained by injection learning strategy which

allows the model to learn from its own error and reduce exposure of
bias, unlike the existing Attention mechanisms that are usually trained
by teacher forcing techniques.
• Different from the existing Attention-based encoder-decoder model that

uses the last hidden state of the encoder as an initial hidden state of the
decoder, the proposed model uses independent and randomly initialized
hidden states for both encoder and decoder layers.
1. Attention-based Encoder-Decoder Model for Amharic OCR 91
• The proposed model is designed by leveraging the architecture of the

Seq2Seq framework, used in NMT, and then stacking CNN layers be-
fore the encoder network as a feature extractor. During training, the
overall model components are treated as a unified framework.
7.1.2 Related Work
The exiting OCR models can utilize either traditional or holistic techniques.
Methods belong to the first category were mainly applied before the introduc-
tion of deep learning, and follows step-wise routines. In contrast, the holistic
approach integrated the feature extraction and sequential translation steps in
a unified framework that are trained from end-to-end. The details are given in
chapter 2. Therefore, in this section, we only review the existing state-of-the-
art techniques that are applied for sequence-to-sequence learning tasks.
The main challenge in sequence-to-sequence learning tasks is to find an

appropriate alignment between input and output sequences of variable length.
For text-image recognition, one needs to identify the correct character at each
time step without any prior knowledge about the alignment between the in-
put image pixels and the target characters. The current two major methods
that overcome this problem are CTC- based and Attention-based sequence-to-
sequence approaches. CTC-based models compute a probability distribution
over all possible output sequences, given an input sequence. They do so by
dividing the input sequence into frames and emitting, for each frame, the like-
lihood of each character of the target alphabet. The probability distribution
can be used to infer the actual output greedily, either by taking the most likely
character at each time step.
In literature, most CTC-based architectures for text-image recognition were

LSTM networks, in many cases using the Bi-LSTM [BHL+ 19a]. Others, like
[LNNN17, BHM+ 20], integrate CNNs for an improved low-level feature ex-
traction prior to the recurrent layers. This approach is applied to printed text-
image recognition [GVK19], handwritten recognition [YBJL12], and license
plate recognition [STA+ 18].
Recently, an alternative idea of sequence-to-sequence architectures that fol-

low the encoder-decoder framework, is to decouple the decoding from the
feature extraction. The models consist of an encoder module, at the bottom
layer, that reads and builds a feature representation of the input sequence,
and a decoder module, at the top layer, which generates the output sequence
one token at a time. The decoder uses the Attention mechanism to gather
context information and search for relevant parts of the encoded features.
Bahdanua [BCB14] and Luong [LPM15] proposed Attention-based encoder-
decoder model for machine language translation. Other works, like Doetsch
et al. [DZN16], apply an Attention neural network to the the output of a
BLSTM network operating on frames extracted from a text line image with a
sliding window and Bluche [Blu16] propose a similar technique for end-to-
end handwritten paragraph recognition.
Finally, our system is based on Bahdanua [BCB14] Attention network ar-
chitecture the one proposed for neural machine translation. The main differ-
ence is that we apply the Attention for text-line image recognition where the
decoder network outputs character by character given the decoder history and
the expected input from the Attention mechanism. Further, CNN layers are
integrated with the encoder network as feature descriptors and unlike Bah-
danau’s networks that use GRU, both the encoder and decoder networks are
LSTM networks. In addition, the hidden state of decoder LSTM is randomly
initialized with a Keras default weight initializer, Xavier uniform initializer,
instead of the final state of the encoder LSTM network. In the case of Bah-
danua’s Attention training was done using the teacher forcing technique while
our proposed model is trained by re-injecting the previous decoder’s predic-
tions into the current decoder’s input.
7.1.3 The Proposed Attention-based Network
In this section, we elaborate on the proposed Attention-based encoder-decoder

network for Amharic OCR. In the text-image recognition task, the raw data
is 32 by 128 text-line images which should be processed and encoded to a
sequence of high-level features. To do so, as illustrated in Figure 7.1, our
Amharic OCR model follows the standard encoder-decoder framework with
the Attention mechanism. The model consists of three basic modules: an en-
coder that combines a CNN as a generic feature extractor with recurrent layers
to introduce temporal contexts in the feature representation, a decoder that uti-
lizes a recurrent layer to interpret those features, and an Attention mechanism
that enables the decoder to focus on the most relevant encoded features at
each decoding time step.
Since our encoder starts with CNNs, the segmented text-line images are
converted into a sequence of visual feature vectors. CNN layers integrated
in encoder network are pretrained model weights adopted from an Amharic
text-image recognizer proposed by Belay [BHM+ 20] (see section 6.2). Then,
Bidirectional-LSTM layers, which read the sequence of convolutional fea-
tures to encode temporal context between them, are employed. The Bidirectional-
LSTM processes the sequence in opposite directions to encode both forward
and backward dependencies and capture the natural relationship of texts.
Suppose that the input text-line image I consists of a sequence length T x,

then the encoder-CNN processes an input image I and transfers it into an
intermediate-level feature map X, which can be thought of as a sequence of
column vectors X =(x1 , x2 , ..., xT x ). This sequence is then processed by two
Bidirectional-LSTM layers and we get the combined state sequence of the fi-
nal encoded feature map H =(h1 , h2 , ..., hT x ), of the forward and backward
hidden states. The third component of this model is a unidirectional LSTM
layer, called the decoder that generates the target character sequence present
in the image given the current image summary and state vector. At each
time-step t, the decoder computes a probability distribution over the possi-
ble characters and predicts the most probable character yt , conditioned on its
own previous predictions (y0 , y1 , ..., yt−l ) and a time-dependent context vector
ct , which contains information from the encoded features. Formally, for an
output sequence length T y, it defines a probability over the output sequence
Y =(y1 , y2 , ..., yT y ) by modeling each condition as p(yt |(y0 , y1 , ..., yTy−1 , ct ))
= sof tmax(f (yt−1 , st−1 , ct )) where f and st−1 represents the current and
previous LSTM hidden states respectively.
At each step of decoding, Attention layer is introduced to focus on the most

relevant part of the encoded feature representation. For each particular input
sequence, attention mechanism has the power to modify the context vector at
each time step based on some similarity of the decoder hidden state sj−1 with
the encoded features h of the context vector cj where cj is computed for each
target character yj as given by Equation (7.1)
Tx
X
cj = wj,i hi (7.1)
i=1
where the wj,i is an attention weight which is computed by soft-maxing its

corresponding alignment score aj,i using Equation (7.2),
exp(aj,i )
wj,i = PT x (7.2)
k=1 exp(ajk )
where aj,i is the alignment score of concatenated annotation between sj−1 and
y0 y1 y2 yTy-1 yTy
..
.
Softmax
Softmax Softmax Softmax
RepeatVector
Decoder
RepeatVector RepeatVector RepeatVector
s1 s2 sTy-1
LSTM LSTM LSTM
..
.
Concat Concat Concat
..
.
s0
RepeatVector
RepeatVector
..
.
Attention layer
RepeatVector
... ... ...

softmax softmax softmax
..
.
FC2 FC2 ... FC2 FC2 FC2 ... FC2 FC2 FC2 ... FC2
FC1 FC1 ... FC1 FC1 FC1 ... FC1 FC1 FC1 ... FC1
..
.
concat concat concat concat concat concat concat concat concat
..
.
h1 h2 hTx
Encoder
LSTM LSTM LSTM

..
.
LSTM LSTM LSTM

..
.
X1 X2 XTx
..
.
CNN-layers
Amharic text-image
Figure 7.1: Attention-based Encoder-Decoder model for Amharic OCR.

The encoder converts an input text-line image I into a sequence
of constant feature vectors h. The decoder generates the output
sequence y =(y0 , y1 , ..., yT y ) one character at a time, where y0 is
the dummy sequence which is generated before the actual target
character sequence started to generate. At each time step t, it uses
an attention mechanism to produce a context vector ct based on
the encoded feature vectors and a time-dependent decoder hidden
state st . This context vector is used to generate the decoders out-
put yt for the current time step. A concatenation of the context
vector at t time step and the output y at t-1 time step serves as
the decoders next input. During concatenation, to have the same
dimension, feature vectors are regenerated using, RepeatVector, a
keras function.
hi at each time step t and it can be computed using Equation (7.3).
aj,i = f (g(hi , sj−1 )), for i = 1, ..., Tx (7.3)
The function g and f in Equation (7.3) are feed-forward neural networks, with
tanh activation function, that are stacked consecutively. The intuition of f and
g is to let the model learn the alignment weights together with the translation
while training the whole model layers.
We carried out the experiments on the ADOCR Amharic database which is

described in detail in chapter 4. The encoder and decoder of the RNNs have
128 hidden units each. The encoder of the RNN consists of forward and back-
ward RNNs each having 128 hidden units. Its decoder also has 128 hidden
units. In both cases, we use a multilayer perceptron network, two in the en-
coder part and one in the decoder part, with a single softmax hidden layer to
compute the conditional probability of each target character. The whole archi-
tecture, depicted in Figure 7.1, computes a fully differentiable function, which
parameters can be trained from end-to-end in a supervised manner via the
backpropagation algorithm. The optimized cost is the negative log-likelihood
of the correct transcription and given as,
L
X
Loss(I,y) = − log(yt |I) (7.4)
t
where I is the image, y = y1 , y2 , ..., yT y is the target character sequence and

p(yt |I) are the output probability of the network.
LSTM networks are employed for both the decoder and encoder modules.
The Attention mechanism, in our model, computes the attention weight and
outputs the context vector once it took all hidden states of the encoder LSTM
and the previous hidden state of the decoder-LSTM. In the entire model, the
Attention module is iteratively called, till it reaches the maximum target se-
quence length, T y times to give back the computed context vector which is
going to be used by the decoder-LSTM. At this time all yt copies have the
same weights by implementing layers with shareable weights. The networks
are trained, to optimize an entropy-based loss, with RMSProp optimizer and
a batch size of 128 for 25 epochs.
The performance of the model is measured using the CER and 1.17% and
GT Input text-line Images Prediction
በኩል የተለመደ ትብብር በኩል የተለመደ ትብር
ዳይሬክተር ለኮምፒዩቲንግ ፋኩልቲ ዳይሬክት ለኮምፒቲንንግ ፋኩልቲ
በአሁኑ ጊዜ ህንፃውንም በአሁኑ ጊዜ ህንዳውን
ዕውን ለማድርግ ሌት ተቀን ዕውን ለማድርግ ሌት ተተቀን
እንዲያገኝ ምቹ ሁኔታን እንዲያገኝ ምቹ ሁኔታን

Figure 7.2: Sample predicted text images using attention-based OCR
model. Text-line images (middle), their corresponding GT texts
(lef t) and model predictions (right).
5.21% on ADOCR test datasets that are synthetically generated with Visual
Geez and Power Geez fonts respectively. We also carried out other experi-
ments on a printed Amharic text-line images dataset from the ADOCR test
dataset and achieved a promising result with a CER of 2.54%. Sample text-
line images that are wrongly recognized during evaluation of our Amharic
OCR model are depicted in Figure 7.2. Characters marked by colored-boxes
are wrong predictions (it can be deletion, substitution or insertion errors). For
example, characters marked by blue-boxes, such as b and r from the first
and second text-line image respectively are sample deleted characters. Other
characters, such as n and °, from second and fourth predicted texts that
are marked with green-boxes are insertion errors, while character × in the
2. Blended Attention-CTC for Amharic Text-image Recognition 97
third predicted text which is marked by red-box is one of the substitution error
recorded during experimentation.
The other type of error that usually affects the recognition performance
of our model is the missing of characters either on the ground-truth or on
the text-line image itself. For example, the character m, marked by violet-
box, in the third text-line image’s ground-truth is one of the character missed
during the text-line image generation. Even though all visible characters in the
text-line image are predicted correctly, since the CER is computed between
the ground-truth texts and the predicted texts, such types of errors are still
considered deletion errors. Besides wrongly predicted text-line images, the
fifth text-line image in Figure 7.2, is a sample correctly predicted text-line
image from ADOCR test sets.
7.1.5 Conclusions
This section successfully approaches the Amharic script recognition with an

Attention-based sequence-to-sequence architecture. The proposed model in-
tegrates convolutional feature extractors with recurrent neural networks to
encode both the visual information, as well as the temporal context in the
input image while separate recurrent neural networks are employed to de-
code the actual Amharic character sequence. Overall, we obtain results that
are competitive with the state-of-the-art recognition performance on ADOCR
datasets. As we observed the empirical results, the Attention-based encoder-
decoder model becomes poor when the sequence length increases. In most
cases, the first characters are always correctly predicted while the rest errors
have no patterns; thus it is hard to learn in the initial training stage for longer
input sequences. Such character errors are not observed in the LSTM-CTC
based networks. The next section presents an OCR model which is designed
by blending the concept of attention into the CTC objective function.
7.2 Blended Attention-CTC for Amharic

Text-image Recognition
In chapter 6 of this thesis, CTC-based networks have been proposed for recog-
nition of Amharic text-line image (see section 6.1 and 6.2). In addition, in sec-
tion 7.1 of this chapter, Attention-based encoder-decoder networks have been
also introduced for Amharic OCR. In all attempts, the recognition perfor-
mance of the proposed approaches is evaluated with the ADOCR test database
and promising experimental results are reported. In this section, we present
a novel Attention-based approach which is designed by blending the concept
of Attention mechanism directly within the CTC objective function. The pro-
posed model consists of an encoder module, Attention module, and transcrip-
tion module in a unified framework. The new approach and implementation
details are presented as follows. Section 7.2.1 talks about related works. The
proposed blended Attention-CTC model is presented in section 7.2.2, and
section 7.2.3 presents the detail of datasets. In the last two subsections, ex-
perimental results and conclusions are presented respectively.
7.2.1 Related work
Even though previous research on Amharic OCR was focused on segmentation-

based OCR models [MJ07, BHL+ 19b, CH03, BHS18], recently published
works [ALT18] and [BHL+ 19a] proposed a Bidirectional LSTM network ar-
chitecture with CTC for Amharic text-line image recognition. End-to-end
learning, that uses CNN, LSTM, and CTC in a unified framework [BHM+ 20],
is also proposed for Amharic OCR and achieved a better recognition perfor-
mance. These segmentation-free attempts made for Amharic script recog-
nition push the research for the OCR of Amharic a step forward. Several
segmentation-free OCR techniques, based on Attention and CTC networks,
have been studied and demonstrated groundbreaking performances for mul-
tiple scrips such as Convolutional Recurrent Neural Network (CRNN) for
Japanese handwritten recognition [LNNN17], segmentation free Chinese hand-
written text recognition [ML15], a hybrid Convolutional-LSTM for text image
recognition [Bre17], Multidimensional LSTM for Chinese handwritten recog-
nition [WYCL17], combined Connectionist Temporal Classification (CTC)
with Bidirectional LSTM for unconstrained online handwriting recognition
[GLB+ 08].
As described in section 7.1, researchers have recently applied the Atten-

tion mechanism in different research areas including in the field of OCR.
Therefore, it becomes popular and a choice of many researchers in the area
of OCR. Attention mechanism has been applied for recognizing handwritten
texts [PV17, CV18], characters in the wild [LO16], handwritten mathemat-
ical expression [ZDD18]. The attention mechanism assists the network in
learning the correct alignment between the input image pixels and the target
characters. In addition, it improves the ability of the network in extracting

the most relevant feature for each part of the output sequence. Inspired by
the success of the Attention mechanism in sequence-to-sequence translation,
we continue to focus on OCR tasks, and we integrate the capability of the
Attention mechanism in the CTC network so as utilize the benefits from both
techniques.
7.2.2 The Proposed Blended Attention-CTC Model
The proposed blended Attention-CTC model, as shown in Figure 7.3, consists

of three modules. The encoder module takes the input features x and maps
them to a higher-level feature representation henc = (h1 enc , h2 enc , ..., hTx enc ).
The Attention module takes the output features henc of the encoder module
and computes the context vector from each hidden features. The output of the
Attention module, attention context information, is passed to the Soft-max
layer in order to produce a probability distribution, P (y 0 , y 1 , ..., y t-1 |x) over
the given input sequence x. Finally, the probability distribution P goes to the
CTC-decoder for transcription.
The proposed model doesn’t change the usual training procedure of CTC
explained in [GLB+ 08]. However, we embed the Attention mechanism, hereby
further described as follows, in between the LSTM and CTC layers. The in-
tuition of the attention layer, in our proposed model, is to produce a weighted
context vector (C vec ) with an improved hidden layer representation, and this
context vector is computed using Equation (7.5).
Tx
X
t
Cvec = sti hi enc (7.5)
i=1
where the st is an attention weight of each annotation hi enc computed by soft-

maxing its corresponding attention score using Equation (7.6),
exp(ai )
si = PT x (7.6)
k=1 exp(ak )
where ai is the attention score of henc
i at each time step t and it can be com-
puted using Equation (7.7).
ai = f (hi enc ), for i = 1, ..., Tx (7.7)
The function f in Equation (7.7) is a feed-forward neural network scoring

የአያቴን ታሪኮች ይዤላችሁ

Output text
CTC-decoder
p(πt |x)
Softmax-2
t enc
Cvec = ∑ st h
i
i=1
s1 s2 ... sT x Attention layer
softmax-1
a1 a2
... aT
x
FC FC ... FC
enc enc enc

h h h
...
1 2 Tx
h0
Encoder-LSTM Encoder-LSTM Encoder-LSTM h0
x
1
x
2 ... T
x
CNN-layers
Input text-image
Figure 7.3: The proposed blended Attention-CTC model. A score (ai ) of

encoder hidden state (henc
i ) at each time-step is computed using
a scoring function described in Equation (7.7), just by propa-
gating the henc
i through fully connected network (FC). Attention
distribution, called attention weights (si ) are computed by run-
ning all scores over a soft-max (Soft-max-1) layer. To compute
the alignment vector , we multiply each henci with its correspond-
ing soft-maxed score (si ). Then, the sum of all alignment vectors
produced the context vector (C vec ), which is an aggregated infor-
mation. Once the C vec is obtained, it passes through the second
soft-max layer ( Soft-max-2) for probability distribution over the n
possible characters in the ground-truth (GT). The output of Soft-
max-2 is a sequence of T time steps of (n + 1) characters which is
then decoded using CTC-decoder.
function that lets the model learn the alignment weights together with the
translation while training the whole model layers.
During training the blended Attention-CTC model, a CTC loss function

(lCTC ) is used to train the network from end-to-end (See Equation (2.5)) and
a target label in path π is obtained by mapping reduction function B, using
the example explained in [BHL+ 19a], that convert a sequence of Soft-max
output for each frame to a label sequence by removing repeated labels and
blank tokens from the sequences of character (C) with the highest score (i)
generated using Equation (6.1).
7.2.3 Database
To train and evaluate the performance of the proposed model, we use the
ADOCR database introduced by [BHL+ 19a]. The original Amharic OCR
database was composed of 337,337 Amharic text-line images, each with mul-
tiple word instances collected from different sources. Sample Amharic text-
line images from ADOCR database and the detail descriptions are given in
chapter 4, section 4.3.2.
7.2.4 Experiments
Our model is trained with the ADOCR database. Similar to section 6.2, im-
ages are scaled to 32 by 128 pixels so as to minimize computations. Since
there is no explicitly stated validation data, in the ADOCR dataset, the same
selection mechanism is applied and randomly selected 7% of the training sam-
ples as validation samples. A blended Attention-CTC network is formulated
by directly taking the advantage of Attention mechanism and CTC network
as an integrated framework and trained in an end-to-end fashion.
During training this model, we use two bi-directional LSTM, each with 128
hidden units and a dropout rate of 0.25, on top of seven convolutional layers
that are stacked serially with ReLU activation function as an encoder. The de-
coder LSTM, presented in section 7.1, is removed and then the CTC objective
function that is blended with the Attention mechanism is in place. The con-
volutional layers of the encoder module are composed of seven convolutional
layers which are already trained in section 6.2 and the same weights are used
in this section via transfer learning.
a c
b d
Figure 7.4: The training & validation losses of model trained with dif-
ferent network settings. (a). CTC loss with an LSTM-CTC
model. (b) CTC loss with a CNN-LSTM-CTC model. (c) CE loss
of Attention-based encoder-decoder model. (d) CTC loss of the
proposed blended Attention-CTC model
We use a batch size of 128 with Adam optimizer and trained for 15 epochs.
Once the probability of labels is obtained from the trained model, we use
best path decoding [Gra08] to generate a character (C i ) that has the maxi-
mum score at each time step t (See Equation (6.1)). The learning loss of the
proposed model and other models trained using different network settings are
depicted in Figure 7.4.
7.2.5 Results
The performance of the proposed blended Attention-CTC model is evaluated

against three test datasets, from the ADOCR database, and compared with
state-of-the-art approaches and the proposed model improves the recognition
performance by 0.67–7.50% from the LSTM-CTC based model and by 2%
from the recently published work with CNN-LSTM-CTC network which use
Figure 7.5: Sample text-line image with the corresponding predicted and
ground-truth texts. Characters, in the predicted text, marked
with red rectangle are wrongly predicted by both Attention based
Encoder-decoder (AED) and Blended Attention-CTC Networks
(BACN).
the same test dataset. Sample Amharic text-image from ADOCR test dataset
with the corresponding GT texts and predictions are depicted in Figure 7.5.
7.2.6 Conclusions
In this section, a Blended Attention-CTC network called BACN, for Amharic

text-line image recognition, is presented. BACN consists of a Bidirectional
LSTM, stacked on top of CNN layers, as an encoder and a CTC layer as
a decoder. To enhance the hidden layer feature representation, the Atten-
tion mechanism is embedded between the LSTM and CTC network layers
without changing the CTC objective function and the training process. All
the encoder, Attention, and CTC modules are trained jointly from end-to-
end and a significant improvement is achieved on all the three ADOCR test
datasets compared with state-of-the-art results. Thus, we can conclude that
the blended Attention-CTC network is more effective for Amharic text im-
age recognition than widely used attention-based encoder-decoder and CNN-
LSTM-CTC based networks as well.
7.3 Chapter Summary
In the previous chapter of this thesis, CTC-based networks were employed

for Amharic text-line image recognition and reported baseline results on test
sets of the ADOCR database. This chapter proposed an alternative and effec-
tive text-image recognizer based on the attention mechanism. The proposed
model consists of CNN integrated with Bidirectional-LSTMs as an encoder
and a unidirectional LSTM as a decoder that is unified as a single frame-
work. Then a comparable recognition performance is achieved against the
three test datasets from the ADOCR Amharic database. Unlike the CTC-
based loss, attention-based networks minimized the categorical cross-entropy
loss. In addition, to gain the advantages of both the attention and CTC-based
networks, a new method is proposed by embedding the attention mechanism
into the CTC network without changing the training process in CTC objective
function. In summary, the blended attention-CTC model outperforms all the
state-of-the-art models on the ADOCR test datasets.
Chapter 8
Conclusions and Recommendations
for Future Work
This chapter presents the conclusions of the research work described in the
thesis. The contributions and objectives of the thesis, outlined in chapter 1,
are reviewed and their achievements are addressed. Factors that limit the effi-
ciency of the present work are also presented. The following section presents
the summary of the most relevant conclusions drawn from this thesis and
based on the findings of this thesis work, recommendations are made for fu-
ture research on Amharic script recognition.
8.1 Conclusions
In this thesis, we addressed the problem of Amharic script recognition and it

presents substantial contributions in the field of Optical Character Recogni-
tion (OCR) for printed documents by extending the use of contemporary deep
learning algorithms in this domain. The nature and complexities of Amharic
script have been highlighted, and the potential of the deep learning technique
is exploited because of its success in the field of computer vision and pattern
recognition.
To enable information access to everyone and break language barriers in

the world, there has been an increasing demand for developing workable OCR
systems. This demand has been partially fulfilled over time and there are nu-
merous works presented for multiple scripts. However, Amharic script has
not been exposed to state-of-the-art OCR research methods. For indigenous
105
106 Conclusions and Recommendations for Future Work
scripts in general, only a few efforts are reported to date, which is not enough
to address the complexities. Amharic, whose speakers exceeds 100 million
people around the globe, is one of the most underrepresented scripts in the
field of NLP. Such resource-limited but unique script presents various chal-
lenges towards the development of an efficient OCR system. Some of the
challenges are the use of a large number of characters in the writing system
and the existence of a large set of visually similar characters. In addition, there
are no standards for font type, size, and style which results in the script being
more complex for recognition. There are numerous documents, in Amharic
script, that is collected for centuries. For further analysis of these documents,
digitization is critical. This signifies the importance of the research in this
thesis.
In this thesis, an optimal OCR model for Amharic script recognition is

constructed using state-of-the-art deep learning techniques. Thus, we explore
different approaches for understanding and addressing problems with an em-
phasis on Amharic script. To this end, extensive works have been made so
as to contribute to designing the OCR model for the recognition of Amharic
printed document images.
This thesis consists of different conceptual and technical contributions which

can be seen from two major perspectives. The first contribution is the creation
of an ADOCR database with various versions and types of datasets that have
been developed to train and evaluate the performance of OCR systems of
Amharic script. This database contains character and text-line images. The
character level database is composed of 80,000 synthetically generated basic
Amharic characters while the second dataset is composed of 337,337 text-
line images. The text-line dataset is created considering the writing system in
Amharic script and the texts are collected from different sources. We believe
that the ADOCR database could be used as a baseline database for further
Amharic OCR system development by researchers in the area.
In the second contribution of this thesis, deep learning-based OCR method-

ologies have been assessed for Amharic script recognition. In this thesis, we
present the major challenges in the development of the OCR model for in-
digenous scripts, like Amharic, which have their own alphabets. Accordingly,
these challenges are due to the nature of the script itself and/or approaches
followed by research attempts made, on Amharic script, so far. All previous
attempts were based on classical machine learning techniques and the mod-
els were trained with segmented characters. Further, the number of datasets
was limited in number and they did not include all the possible characters
1. Conclusions 107
used in Amharic script. Moreover, none of these datasets are publicly avail-
able. Hence, we have explored state-of-the-art OCR tools and techniques
ranging from dataset preparation to model development, variants of artificial
neural networks ranging from convolutional neural networks to recurrent net-
works, sequence-to-sequence learning, and alignment mechanisms ranging
from CTC to Attention mechanism. Based on the results from empirical ob-
servations and conceptual frameworks from the theoretical foundation, sev-
eral conclusions can be drawn in this thesis work.
• The shortage of training dataset can be solved with artificial data, how-
ever, it should be generated carefully to reflect scanned real-world doc-
uments as close as possible. Experiments performed, in this thesis, by
training different deep-learning-based OCR models on synthetic data
and testing them on real scanned Amharic documents justifies this claim.
This thesis presents the procedures for synthetic data generation and
real scanned data preparation. Further, the dataset used for Amharic
script recognition, in this thesis work, is now made publicly available
at http://www.dfki.uni-kl.de/˜belay/ for free.
• This thesis presented novel CNN-based Amharic character recognition
techniques. The approach employed a Factored-CNN network that is
designed based on the grapheme of basic Amharic characters called
Fidel-Gebeta and achieved character recognition accuracies that are sig-
nificantly better than the performance of existing Amharic characters
recognition techniques.
• Stacking CNNs, before the LSTM network layers, as a feature extrac-
tor and training the hybrid networks using a CTC objective function in
an end-to-end fashion improves the recognition performance than the
LSTM-CTC-based network model.
• Attention embedded RNNs, from neural machine translation tasks, are
adopted and employed for Amharic text-line image recognition. The
recognition efficiency reached near to the performance of CTC-based
sequence alignments.
• The integration of the Attention mechanism and CTC objective func-
tion makes the sequence-to-sequence learning network perform better,
instead of employing each of them separately.
Experimental results show the effectiveness of our approach for the recogni-
tion of Amharic scripts. The majority of the recognition errors observed are
due to improper annotation among text-line images and the ground truth while
108 Conclusions and Recommendations for Future Work
other errors are due to the existence of deformed and/or missing characters in
the beginning and/or end of the text-line images during data generation from
the given texts. In general, the strategies followed, in this thesis, are promising
for the recognition of large digitized Amharic document images. Therefore,
the present approaches could be extended, redesigned or new approaches need
to be proposed for the OCR of historical and handwritten Amharic documents.
Tasks that are not addressed in this thesis and the possibilities to extend the
current work are provided in the next section.
8.2 Recommendations for Future Work
To advance the OCR of Amharic script, in this thesis, various deep learning-
based schemes are proposed. Based on the findings and progression of the
current work, we recommend the following research directions as a possi-
ble future work so as to design an efficient OCR system and digitization of
Amharic language in general.
• In the present work an attempt is made to develop the OCR model and
introduce a database for recognition of Amharic script. It pushes the
technology of Amharic script one step towards digitization. To reduce
the barriers of this language in the digital world, we need appropriate
data that can be used to analyze the linguistics of Amharic. Neural
machine translation and speech recognition need notable attention in
the future since they have not yet been sufficiently investigated.
• The database introduced in this thesis considers only the two commonly
used fonts of Amharic script. Even though a large Amharic database is
created with those fonts, there are other fonts rarely used in Amharic
writing system that is not included in the training set, which may af-
fect the recognition performance of the OCR model. To facilitate OCR
researches and develop versatile OCR for Amharic script, we need to
develop a large database of Amharic texts with different font types and
styles.
• The OCR models proposed in this thesis are based on various state-of-
the-art deep learning-based techniques. The recognition performance
of these models is evaluated using a test dataset from the ADOCR
database. In this regard, we have achieved promising results. Compared
to the over centuries Amharic document collection, which contains im-
ages, tables, unknown fonts, and non-standard scripting, the ADOCR
2. Recommendations for Future Work 109
database is not complex enough. We believe that the current model

should fit for recognition of data with more complex layouts and it could
also be explored for scene Amharic text-image recognition hence it is
recommended as a future research direction.
• Training and evaluation of the OCR methodology, in this thesis, are
based on synthetic and scanned Amharic documents. However, the
same approach can be extended for the case of documents captured with
other document digitizing technology like a camera. Digitizing docu-
ments with the camera may introduce several issues hence as future
work we need to integrate some document preprocessing techniques so
as to correct this issue and simplify the task of recognition with the
current model.
• In this research an OCR system is developed for synthetically generated
and printed Amharic documents. Therefore, it is important to extend the
current research work for the recognition of handwritten and historical
Amharic documents written on parchments and vellums.
Finally, we believe that the approaches followed in this thesis successfully
fulfill the gaps that were presented in Amharic script recognition. We hope it
would also serve as a stepping stone for the future researcher who will work
in the OCR domain specifically for such indigenous and low-resource scripts.
Appendices
111
Appendix A
A.1 Attention Mechanism
This part of the thesis is an appendix that presents the details of the Atten-
tion mechanism and internal working of Attention techniques. Various ar-
chitectures of encoder-decoder networks that have been used commonly for
sequence-to-sequence tasks, including NMT and OCR are also described in
detail. Besides, it gives the necessary mathematical details of Attention-based
networks and basic parameters of attention mechanism that one needs to tune
when training the attention-based encoder-decoder networks.
The concept of Attention in the neural process has been largely studied
in Neuroscience and Computational Neuroscience first [MZ17, IKN98] and
consequently, there were many attempts to replicate it in Machine Learning
[SVL14, LPM15]. In such fields, a particular aspect of the study is visual
attention; thus many animals focus on specific parts of their visual inputs
to compute adequate responses. This principle has a large impact on neural
computation as we need to select the most pertinent piece of information,
rather than using all available information, a large part of it is irrelevant to
compute the neural response. A neural network is considered to be an effort
to mimic human brain actions in a simplified manner. Attention Mechanism
is also an attempt to implement the same action of selectively concentrating
on a few relevant things while ignoring others in deep neural networks.
A similar idea, focusing on specific parts of the input, has been followed in
deep learning, for speech recognition, neural machine translation, visual ques-
tion and answering, and text-image recognition. In the recent trends of deep
learning Attention mechanisms are one of the most exciting advancements,
and that they are here to stay.
113
114
Attention Model
S0
h1 h2 ... hn
Figure A.1: General setting of Attention model. The initial state s0 is the
beginning of the generated sequence, the hi are the representa-
tions of the parts of the input sequence, and the output is a repre-
sentation of the filtered sequence, with a filter putting the focus of
the interesting part for the sequence currently generated.
A.1.1 Generic Model of Attention Mechanism
As illustrated in Figure A.1 the generic setting of Attention model takes a se-
quence with a length of n arguments h1 , ..., hn (in the preceding examples,
the hi would be the hidden states of the encoder network), and a state s0 . It
return a vector C which is supposed to be the summary of the hi , focusing on
information linked to the state s0 . More formally, it returns weighted arith-
metic mean of the hi , and the weights are chosen according to the relevance
of each hi given the state s0 .
The Attention model depicted in Figure A.1 looks a black-box and it doesn’t
clearly show how the summarized information C is computed; thus the whole
structure of the Attention model need to be discussed in detail and the model
should be redrawn by expanding the black-box design.
The intuition of mathematical computation, in the Attention model, is to

produce a vector called context vector c; thus the network computes f1 , f2 , ..., fn
with a tanh activation layer. fi is the aggregated value of hi and s0 , where
each fi is computed, independently, without looking at the other hj for j is
1. Attention Mechanism 115
not equal to i and it is given as,
fi = (Ws0 f s0 + Whs0 hi ) (A.1)
where W is weight matrices to be learned, later, while training the whole

model layers. Then each weight metrics of fi is computed using Softmax
function which is also given as,

exp(fi ))
sof tmax(f1 , ..., fn ) = Pn i, for i = 1, ..., n (A.2)
k=1 exp(fk )
where the si are the softmax of the fi projected on a learned direction and
Pn
i=1 si = 1. The output c is the weighted arithmetic mean of all the hi ,
where the weight represents the relevance for each variable according to the
initial hidden state s0 and computed as,
n
X
c= hi si (A.3)
i=1
A.1.2 Types of Attention and Seq2seq Model
The motivation behind the Attention mechanism is the problem of forget-

ting long source sequences during translation. Seq2seq model aims to trans-
form an input sequence to a target one and both sequences can be of arbitrary
lengths [SVL14]. Examples of transformation tasks include machine transla-
tion between multiple languages in either text or audio, question-answer dia-
log generation, or even image-to-text conversion in OCR. The seq2seq model
normally has an encoder-decoder architecture, composed of two processes.
• Encoding processes: It takes a variable-length input sequence and com-

presses the information into a context vector of a fixed length. This
representation is expected to be a good summary of the meaning of the
whole input sequences.
• Decoding process: It is initialized with the context vector to emit the

transformed output sequences. In the seq2seq model, the encoder and
the decoder networks are linked with the hidden state information which
is sent from the last hidden state of the encoder network to the decoder
as an initial hidden state of the decoder network.
In seq2seq architecture both the encoder and decoder are recurrent neu-
ral networks, which are usually either LSTM or GRU units [CVMG+ 14,
116
መማር አዲስ መረዳትን የማግኘት ሂደት ነው <EOS>
DECODER
(h0,c0) (h1,c1) (h2,c2) (h3,c3) (h4,c4) (h5,c5) (h6,c6) (h7,c7)
<SOS> መማር አዲስ መረዳትን የማግኘት ሂደት ነው

ENCODER
(h0,c0) (h1,c1) (h2,c2) (h3,c3) (h4,c4) (h5,c5) (h6,c6) (h7,c7) (h8,c8)
learning is the process of gaining new understanding
Figure A.2: Seq2seq model. This encoder-decoder model, translating the

sentence ”Learning is the process of gaining new understanding”
to its equivalent Amharic sentence
BCB14]. An example of seq2seq model illustrated in Figure A.2, where

the encoder reads a sequence input with variable lengths, e.g., English words
(Learning is the process of gaining new understanding), and the decoder
produces a sequence output, e.g., corresponding Amharic words, (Œ≈r €Ös
Œ×tn Î√Úxt ‚Ôt ¶Ã). As shown in Figure A.2, the encoder net-
work (an LSTM) takes a sequence of an English word as input and encap-
sulates the information as the internal state vectors which are later used by
the decoder. Since LSTM takes only one element at a time, so if the input
sequence is of length T, then LSTM takes T time steps to read the entire se-
quence (i.e in this example the English sentence has 8 words, then the LSTM
will read this sequence word by word in 8 time-steps ).
The hidden state and cell state (h0 and c0 ) of the encoder are initialized ran-
domly with a smaller value or it can be zero. The dimensions of both states
should be the same as the number of units in the LSTM cell. The encoder will
read the English sentence word by word and store the final internal states, an
intermediate vector, of the LSTM generated after the last time step and since
the output will be generated once the entire sequence is read, therefore out-
puts of the Encoder at each time step have no use and so they are discarded.
The final states of the encode (i.e in this case h8 and c8 ) are the relevant in-
formation of the whole English sentence ”Learning is the process of gaining
new understanding”.
The decoder works in a very similar fashion as the encoder. However, dur-
ing the training and testing phase, the decoder behaves differently unlike the
encoder which is the same in both cases. The start and end of the sequence,
for the decoder, are marked with < SOS > and < EOS > respectively.
During training the initial states (h0 and c0 ) of the decoder is set to the final
states of the encoder (h8 and c8 ). The decoder is trained to generate the out-
put based on the information gathered by the encoder by taking < SOS >
as the first input to generate the first Amharic word and at last, the decoder
learns to predict the, < EOS >. The input at each time step is actual out-
put which is usually called the teacher forcing technique. However, the loss
is calculated on the predicted outputs from each time step and the errors are
backpropagated through time to update the parameters of the model.
Unlike the training phase which uses the actual output as input at each time
step, the test phase uses the protected output produced at each time step as
input in the next time step. The final states of the decoder are discarded as we
got the output hence it is of no use in both the training and testing stages.
A critical and apparent disadvantage of a fixed-length context vector, cre-
ated in the encoding process of the seq2seq model, is the incapability of re-
membering long sentences. Often it has forgotten the first part once it com-
pletes processing the whole input; thus Attention mechanism was introduced
to solve this issue. Even though the first model of Attention mechanisms is
introduced by Bahdanau [BCB14], nowadays, there are various types of At-
tention mechanisms proposed by many researchers from the deep learning
community. The underlying principles of Attention are the same in these all
types of Attention mechanisms that were proposed so far, the differences lie
mainly in their architectures and computations.
Attention can be general and self-attention, hard and soft-attention, global
and local-attention. General-attention manages and quantifying the inter-
dependence between the input and output elements [BCB14], where self-
attention and interdependency within the input elements and relating different
positions of a single sequence to compute a representation of the same se-
quence [CDL16].
The type of Attention that attends to the entire input state space and uses
all the encoder hidden states is also known as global-attention. In contrast,
local-attention attends the part of the input sequence or patch of an image and
uses only a subset of the encoder hidden states [LPM15, XBK+ 15]. Consid-
ering the neural machine translation tasks, the global-attention has a draw-
back that it has to attend to all words on the source side for each target
word, which is expensive and can potentially render it impractical to trans-
118
Decoder
...
GRU
Initial Decoder
hidden states
concat
context vector
+
X X X
...
SoftMax
Attention layer
score score ... score
Decoder
hidden state
Encoder
... hidden state
...
Encoder
GRU
Initial encoder
hidden states
Input#1 Input#2 Input#n
Figure A.3: A typical architecture of Soft attention .
late longer sequences such as paragraphs or documents; thus, to address this

deficiency, a local-attentional mechanism is proposed that can choose to focus
only on a small subset of the source positions per target word. Besides, soft-
attention refers to the global-attention [LPM15] approach in which weights
are placed softly and attends overall patches in the input sequence. The hard-
attention [XBK+ 15], on the other hand, selects one patch of the image/part
of the input sequence to attend at a time. While less expensive at inference
time, the hard-attention model is non-differentiable and requires more compli-
cated techniques such as variance reduction or reinforcement learning to train
[XBK+ 15]. In this case, the local-attention mechanism proposed by Luong
[LPM15] is selectively focused on a small window of context and is differen-
tiable and it avoids the expensive computation incurred in the soft-attention
and at the same time, is easier to train than the hard-attention approach pro-
posed by Xu [XBK+ 15].
Further, the Attention mechanism proposed by Bahdanau [BCB14] is also

a soft-attention and differentiable as well. However, it is expensive in longer
sequence inputs. Global-attention is somehow similar to soft-attention while
the local one is a blend between hard and soft-attention of Xu[XBK+ 15] and
improvement over the hard-attention to make it differentiable. Therefore, At-
tention mechanisms generally can be classified as Global/soft as presented in
the work of Xu [XBK+ 15] and Bahdhnau [BCB14] or Local/Hard attention as
proposed by Xu [XBK+ 15] and Luong [LPM15]. The basic architectural dif-
ferences and working procedures of soft and hard attention, based on the work
of Bahidanu and Luong are illustrated in Figure A.3 and A.4 respectively.
In soft attention, [BCB14], the encoder is a bidirectional gated recurrent

unit (BiGRU) while the decoder is a GRU whose initial hidden state is a vec-
tor modified from the last hidden state of the backward encoder GRU. The
general steps followed in this Attention mechanism can be presented as fol-
lows:
1. Initialize hidden states of the encoder network: The initial encoder

hidden states are a random small number or zero and then the encoder
initialize hidden states of each element in the input sequence.
2. Initialize the decoder hidden state: The initial hidden state of the de-
coder network is a modified vector of the last hidden state of backward
encoder GRU or we can also initialize the decoder hidden states ran-
domly as we did in the encoder network.
3. Compute alignment scores: Alignment score is calculated between the

previous decoder’s hidden state at t − 1 and each of the encoder’s hid-
den states at time-steps t. In this Attention mechanism, the last encoder
hidden state can be used as the initial hidden state of the decoder, and
the alignment score function used in this Attention model is called ad-
ditive or concatenation (i.e Since the decoder hidden state and encoder
outputs will be passed through their linear layer and have their train-
able weights, the current encoder hidden state and the previous decoder
hidden states are first added/concatenated and then tanh is applied).
Alignment score computation is given as,
salign = Wshared .tanh(W enc · henc + W dec · hdec ) (A.4)
4. Softmaxing the alignment scores: Each computed alignment score

runs through a softmax layer.
120
5. Computing the Context Vector: The encoder hidden states and their
respective soft-maxed alignment scores are multiplied and then summed
up the results to form the context vector.
6. Decoding the Output: the context vector is concatenated with the pre-
vious decoder output and fed into the Decoder for that time step along
with the previous decoder hidden state to produce a new output. The
process from steps 3 to 6 is repeated for each time step of the decoder
until completing the specified maximum length of the output.
The second type of Attention is based on the work of Luong [LPM15]

and is often referred to as local/ Multiplicative Attention. In this attention
mechanism, the encoder is a two-stacked long short-term memory (LSTM)
network while the decoder also has the same architecture, whose initial hidden
states are the last hidden states of the encoder network. The input sequences
are feed to the encoder in reverse order. Some of the major steps, based on the
work of Luong [LPM15], followed in this attention are described as follows.
1. Initialize the encoder hidden states: As done in the work of Bah-

danau, hidden states of the encoder unit are initialized randomly.
2. Initialize the decoder hidden states: Once the initial hidden states
of the encoder unit are produced then the last hidden states of both
encoder LSTM are passed to the decoder unit as initial hidden states of
the decoder.
3. Decoder RNN configuration: Since it has two LSTM network layers,

the previous decoder RNN hidden state and decoder output is passed
through the second decoder RNN to generate a new hidden state for
that time step.
4. Compute alignment scores: Using the new decoder hidden state and
the encoder hidden states, alignment scores are computed. In this case,
the Luong Attention uses three types of scoring functions that use the
encoder outputs and the decoder hidden state produced in the previous
step to calculate the alignment scores. The computation for each of the
scoring function is give as follows:
• Dot: In this scoring function the alignment score salign is com-

puted by taking the hidden states of the encoder henc and multiply
them with the hidden state of the decoder hdec . it is given as,
salign = henc · hdec (A.5)

Decoder
LSTM
FC
Initial Decoder
hidden states ...
from encoder
concat
context vector
+
X X X
...
SoftMax
Attention layer
score score ... score
Decoder
hidden state
Encoder
hidden state
...
Encoder
LSTM
Initial encoder
hidden states
...
Input#1 Input#2 Input#n
Figure A.4: Attention model architecture based on the work of Luong.
• General: This scoring function is similar to the dot function, ex-

cept that a weight matrix is applied on their product and it is give
as,
salign = W (henc · hdec ) (A.6)
• Concat: The function Concat scoring function of Luong’s At-

tention is slightly similar to Bahdanau’s one, whereby the de-
coder hidden state is added to the encoder hidden states. How-
ever, In Luong’s Attention the decoder hidden state and encoder
hidden states have shared weights since both of them are added
together first before being passed through a Linear layer and then
this weight matrix will be applied on the tanh activated outputs.
122
It can be computed as,
salign = W.tanh(Wshared (henc + hdec )) (A.7)
5. Softmaxing the alignment scores: the alignment scores for each en-
coder hidden state are passed through Softmax function.
6. Computing the context vector: To compute the context vector, the en-
coder hidden states and their corresponding alignment scores are mul-
tiplied and then summed up.
7. Producing the output: The concatenated context vector and the de-
coder hidden state generated in step 3 are passed through a feed-forward
neural network to produce a new final output.
Steps 3 to 8 are repeated themselves till the end of the maximum sequence
length.
The Attention mechanism proposed by Luong is different from the one pre-
sented by Bahdanau with some features. For example, attention calculation
in Bahdanau requires the output of the decoder from the prior time step while
Luong Attention uses the output from the encoder and decoder for the cur-
rent time step only. Besides, Loung uses a reversed input sequence instead of
bidirectional inputs, LSTM instead of GRU elements and unlike Bahdanau’s
Attention, Luong uses the dropout rate.
A.1.3 Basic Parameters of Attention-based Networks
As done in all neural network-based models, there are various network pa-
rameters that one needs to tune and use so that the Attention-based networks
perform better on a given task. In this case, all parameters of the other module
(the encode and decoder modules) should have selected and tuned as usual,
however, in this section, in addition to tunable parameters used in this module,
the configuration of the Attention module are also described.
Parameters for Encoder and Decoder Networks: The number of hidden
layers, the number of LSTM cells, and even the activation function in an
individual layer are the important parameters that need to be properly selected
and tuned. However, we need also consider the effect of employing more
layers. since the learning capability of a network increases as it learns more
compact representation, but it may also make the training more difficult. We
need also properly select the type of Attention ( either soft or hard-Attention)
and training procedure ( either teacher-forcing or injection learning). Finally,

we need to tune the type of scoring function, number of neurons, and size of
fully connected network layers in the Attention layer.
Bibliography
[AB08] Yaregal Assabie and Josef Bigun. Writer-independent offline recognition of

handwritten ethiopic characters. Proc. 11th ICFHR, pages 652–656, 2008. 14
[AB09] Yaregal Assabie and Josef Bigun. Hmm-based handwritten amharic word
recognition with feature concatenation. In 2009 10th International Conference
on Document Analysis and Recognition, pages 961–965. IEEE, 2009. 38, 62, 63
[AB11] Yaregal Assabie and Josef Bigun. Offline handwritten amharic word recogni-
tion. Pattern Recognition Letters, 32(8):1089–1099, 2011. 72
[AEP07] Andreas Argyriou, Theodoros Evgeniou, and Massimiliano Pontil. Multi-task

feature learning. In Advances in neural information processing systems, pages
41–48, 2007. 64
[AHB+ 18] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson,
Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image
captioning and visual question answering. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition, pages 6077–6086, 2018.
27
[Ale97] Worku Alemu. The application of ocr techniques to the amharic script. Master’s
thesis, School of Information Science, Addis Ababa University, Addis Ababa,
1997. 32, 38, 62, 63, 72, 73
[All89] Alan Allport. Visual attention. 1989. 22
[All95] May Allam. Segmentation versus segmentation-free for recognizing arabic text.
In Document Recognition II, volume 2422, pages 228–235. International Soci-
ety for Optics and Photonics, 1995. 17
[ALT18] Direselign Addis, Chuan-Ming Liu, and Van-Dai Ta. Printed ethiopic script
recognition by using lstm networks. In 2018 International Conference on Sys-
tem Science and Engineering (ICSSE), pages 1–6. IEEE, 2018. 38, 87, 98
[AM90] Yaser S Abu-Mostafa. Learning from hints in neural networks. Journal of

complexity, 6(2):192–198, 1990. 64
125
126 Bibliography
[AMKS15] R Anil, K Manjusha, S Sachin Kumar, and KP Soman. Convolutional neural

networks for the recognition of malayalam characters. In Proceedings of the
3rd International Conference on Frontiers of Intelligent Computing: Theory
and Applications (FICTA) 2014, pages 493–500. Springer, 2015. 55
[AMT+ 18] Solomon Teferra Abate, Michael Melese, Martha Yifiru Tachbelie, Million
Meshesha, Solomon Atinafu, Wondwossen Mulugeta, Yaregal Assabie, Hafte
Abera, Binyam Ephrem Seyoum, Tewodros Abebe, et al. Parallel corpora for
bi-lingual english-ethiopian languages statistical machine translation. In Pro-
ceedings of the 27th International Conference on Computational Linguistics,
pages 3102–3111, 2018. 4
[Aro85] Mark Aronoff. Orthography and linguistic theory: The syntactic basis of ma-
soretic hebrew punctuation. Language, pages 28–72, 1985. 32
[Asn12] Biniam Asnake. Retrieval from real-life amharic document images. Master’s
thesis, School of Information Science, Addis Ababa University, 2012. 31
[Ass02] Yaregal Assabie. Optical character recognition of amharic text: an integrated

approach. Master’s thesis, Addis Ababa University, Addis Ababa, 2002. 38, 54,
55, 56, 60, 73
[AT17] Akm Ashiquzzaman and Abdul Kawsar Tushar. Handwritten arabic numeral
recognition using deep learning neural networks. In 2017 IEEE International
Conference on Imaging, Vision & Pattern Recognition (icIVPR), pages 1–4.
IEEE, 2017. 54, 55
[BCB14] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural ma-
chine translation by jointly learning to align and translate. arXiv preprint
arXiv:1409.0473, 2014. 22, 23, 24, 26, 90, 92, 116, 117, 119
[BCFX14] Jinfeng Bai, Zhineng Chen, Bailan Feng, and Bo Xu. Image character recogni-
tion using deep convolutional neural network learned from different languages.
In 2014 IEEE International Conference on Image Processing (ICIP), pages
2560–2564. IEEE, 2014. 54, 55
[BHL+ 19a] B. H. Belay, T. Habtegebrial, M. Liwicki, G. Belay, and D. Stricker. Amharic

text image recognition: Database, algorithm, and analysis. In 2019 Inter-
national Conference on Document Analysis and Recognition (ICDAR), pages
1268–1273. IEEE, Sep. 2019. 3, 43, 86, 87, 90, 91, 98, 101
[BHL+ 19b] Birhanu Belay, Tewodros Habtegebrial, Marcus Liwicki, Gebeyehu Belay, and
Didier Stricker. Factored convolutional neural network for amharic character
image recognition. In 2019 IEEE International Conference on Image Process-
ing (ICIP), pages 2906–2910. IEEE, 2019. 19, 32, 72, 98
[BHM+ 20] Birhanu Belay, Tewodros Habtegebrial, Million Meshesha, Marcus Liwicki,
Gebeyehu Belay, and Didier Stricker. Amharic ocr: An end-to-end learning.
Applied Sciences, 10(3):1117, 2020. 32, 91, 93, 98
Bibliography 127
[BHS18] Birhanu Hailu Belay, Tewodros Amberbir Habtegebrial, and Didier Stricker.
Amharic character image recognition. In 2018 IEEE 18th International Con-
ference on Communication Technology (ICCT), pages 1179–1182. IEEE, 2018.
62, 67, 72, 98
[Blu16] Théodore Bluche. Joint line segmentation and transcription for end-to-end
handwritten paragraph recognition. In Advances in Neural Information Pro-
cessing Systems, pages 838–846, 2016. 92
[Bre08] Thomas M Breuel. The ocropus open source ocr system. In Document Recog-
nition and Retrieval XV, volume 6815, page 68150F. International Society for
Optics and Photonics, 2008. 3, 44, 55, 56
[Bre17] Thomas M Breuel. High performance text recognition using a hybrid

convolutional-lstm implementation. In 2017 14th IAPR International Confer-
ence on Document Analysis and Recognition (ICDAR), volume 1, pages 11–16.
IEEE, 2017. 72, 98
[BSS17] Erik Bochinski, Tobias Senst, and Thomas Sikora. Hyper-parameter optimiza-
tion for convolutional neural network committees based on evolutionary algo-
rithms. In 2017 IEEE International Conference on Image Processing (ICIP),
pages 3924–3928. IEEE, 2017. 19
[BT14] Henry S Baird and Karl Tombre. The evolution of document image analy-
sis. In Handbook of document image processing and recognition, pages 63–71.
Springer, 2014. 1
[BUHAAS13] Thomas M Breuel, Adnan Ul-Hasan, Mayce Ali Al-Azawi, and Faisal Shafait.
High-performance ocr for printed english and fraktur using lstm networks. In
2013 12th International Conference on Document Analysis and Recognition,
pages 683–687. IEEE, 2013. 54, 55
[CAC+ 04] Luke Cole, David Austin, Lance Cole, et al. Visual object recognition using
template matching. In Australian conference on robotics and automation, 2004.
16
[CBC19] SA Chekole, F Bekele, and B Chekol. Archival management and preservation

in ethiopia: The case of east gojjam. J Tourism Hospit, 8(408):2167–0269,
2019. 5
[CCB15] Kyunghyun Cho, Aaron Courville, and Yoshua Bengio. Describing multimedia
content using attention-based encoder-decoder networks. IEEE Transactions
on Multimedia, 17(11):1875–1886, 2015. 90
[CDL16] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-
networks for machine reading. arXiv preprint arXiv:1601.06733, 2016. 117
[CdSBJB+ 06] Paulo Rodrigo Cavalin, Alceu de Souza Britto Jr, Flávio Bortolozzi, Robert
Sabourin, and Luiz E Soares Oliveira. An implicit segmentation-based method
128 Bibliography
for recognition of handwritten strings of characters. In Proceedings of the 2006

ACM symposium on Applied computing, pages 836–840, 2006. 15
[CH03] John Cowell and Fiaz Hussain. Amharic character recognition using a fast
signature based algorithm. In Information Visualization, 2003. IV 2003. Pro-
ceedings. 7th International Conference on, pages 384–389. IEEE, 2003. 14, 72,
98
[Cho14] Amit Choudhary. A review of various character segmentation techniques for

cursive handwritten words recognition. Int. J. Inf. Comput. Technol, 4(6):559–
564, 2014. 14
[CL96] Richard G Casey and Eric Lecolinet. A survey of methods and strategies in
character segmentation. IEEE transactions on pattern analysis and machine
intelligence, 18(7):690–706, 1996. 14
[Cou03] Florian Coulmas. Writing systems: An introduction to their linguistic analysis.

Cambridge University Press, 2003. 31, 32
[CPT+ 19] J. Cai, L. Peng, Y. Tang, C. Liu, and P. Li. Th-gan: Generative adversarial net-
work based transfer learning for historical chinese character recognition. In
2019 International Conference on Document Analysis and Recognition (IC-
DAR), pages 178–183. IEEE, Sep. 2019. 3
[CRA13] Amit Choudhary, Rahul Rishi, and Savita Ahlawat. A new character segmen-
tation approach for off-line cursive handwritten words. Procedia Computer
Science, 17:88–95, 2013. 15
[CV18] Arindam Chowdhury and Lovekesh Vig. An efficient end-to-end neural model
for handwritten text recognition. arXiv preprint arXiv:1807.07965, 2018. 27, 98
[CVMG+ 14] Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau,
Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase repre-
sentations using rnn encoder-decoder for statistical machine translation. arXiv
preprint arXiv:1406.1078, 2014. 116
[DCBC15] Long Duong, Trevor Cohn, Steven Bird, and Paul Cook. Low resource depen-
dency parsing: Cross-lingual parameter sharing in a neural network parser. In
Proceedings of the 53rd Annual Meeting of the Association for Computational
Linguistics and the 7th International Joint Conference on Natural Language
Processing (Volume 2: Short Papers), pages 845–850, 2015. 63
[Dem11] Fitsum Demissie. Developing optical character recoginition for ethiopic scripts,
2011. 60
[DHK13] Li Deng, Geoffrey Hinton, and Brian Kingsbury. New types of deep neural net-
work learning for speech recognition and related applications: An overview. In
2013 IEEE international conference on acoustics, speech and signal process-
ing, pages 8599–8603. IEEE, 2013. 63
Bibliography 129
[DJH01] O De Jeses and Martin T Hagan. Backpropagation through time for a general
class of recurrent network. In IJCNN’01. International Joint Conference on
Neural Networks. Proceedings (Cat. No. 01CH37222), volume 4, pages 2638–
2643. IEEE, 2001. 19
[DLY+ 19] Amit Das, Jinyu Li, Guoli Ye, Rui Zhao, and Yifan Gong. Advancing acoustic-
to-word ctc model with attention and mixed-units. IEEE/ACM Transactions on
Audio, Speech, and Language Processing, 27(12):1880–1892, 2019. 90
[DT+ 14] David Doermann, Karl Tombre, et al. Handbook of document image processing
and recognition. Springer, 2014. 1
[DZN16] Patrick Doetsch, Albert Zeyer, and Hermann Ney. Bidirectional decoder
networks for attention-based end-to-end offline handwriting recognition. In
2016 15th International Conference on Frontiers in Handwriting Recognition
(ICFHR), pages 361–366. IEEE, 2016. 92
[EF20] Gary F. Simons Eberhard, David M. and Charles D. Fennig. Ethnologue: Lan-
guages of the World. SIL International, Dallas, Texas, United States, twenty-
third edition, 2020. 29
[Eik93] Line Eikvil. Optical character recognition. citeseer. ist. psu. edu/142042. html,
1993. 14, 16
[ETK16] Mohamed Elleuch, Najiba Tagougui, and Monji Kherallah. A novel architec-
ture of cnn based on svm classifier for recognising arabic handwritten script.
International Journal of Intelligent Systems Technologies and Applications,
15(4):323–340, 2016. 62
[GCBG08] Emmanuèle Grosicki1 , Matthieu Carre, Jean-Marie Brodin, and Edouard

Geoffrois1 . Rimes evaluation campaign for handwritten mail processing. 2008.
43
[GDS10] Debashis Ghosh, Tulika Dube, and Adamane Shivaprasad. Script recogni-
tion—a review. IEEE Transactions on pattern analysis and machine intelli-
gence, 32(12):2142–2161, 2010. 54, 55
[GFGS06] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhu-
ber. Connectionist temporal classification: labelling unsegmented sequence
data with recurrent neural networks. In Proceedings of the 23rd international
conference on Machine learning, pages 369–376, 2006. 20, 21, 74, 84
[Gla17] Tobias Glasmachers. Limits of end-to-end learning. arXiv preprint

arXiv:1704.08305, 2017. 86
[GLB+ 08] Alex Graves, Marcus Liwicki, Horst Bunke, Jürgen Schmidhuber, and Santi-
ago Fernández. Unconstrained on-line handwriting recognition with recurrent
neural networks. In Advances in neural information processing systems, pages
577–584, 2008. 72, 90, 98, 99
130 Bibliography
[GMH13] Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recog-
nition with deep recurrent neural networks. In 2013 IEEE international con-
ference on acoustics, speech and signal processing, pages 6645–6649. IEEE,
2013. 19
[Gra08] A Graves. Supervised sequence labelling with recurrent neural networks [ph.
d. dissertation]. Technical University of Munich, Germany, 2008. 102
[GSR15] Shivansh Gaur, Siddhant Sonkar, and Partha Pratim Roy. Generation of syn-
thetic training data for handwritten indic script recognition. In 2015 13th Inter-
491–495. IEEE, 2015. 72
[GSTBJ19] Mesay Samuel Gondere, Lars Schmidt-Thieme, Abiot Sinamo Boltena, and
Hadi Samer Jomaa. Handwritten amharic character recognition using a convo-
lutional neural network. arXiv preprint arXiv:1909.12943, 2019. 38, 72
[GVK19] Rajib Ghosh, Chirumavila Vamshi, and Prabhat Kumar. Rnn based online hand-
written word recognition in devanagari and bengali scripts using horizontal zon-
ing. Pattern Recognition, 92:203–218, 2019. 81, 91
[GVZ16] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for
text localisation in natural images. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 2315–2324, 2016. 42
[HA13] Abraham Hailu and Yaregal Assabie. Itemsets-based amharic document cate-
gorization using an extended a priori algorithm. In Language and Technology
Conference, pages 317–326. Springer, 2013. 32
[HA17] Seid Ali Hassen and Yaregal Assabie. Recognition of double sided amharic
braille documents. International Journal of Image, Graphics and Signal Pro-
cessing, 9(4):1, 2017. 62
[Hag15] Tesfahun Hagos. Ocr system for the recognition of ethiopic real life documents.
Master’s thesis, Bahir Dar Institute of Technology, 2015. 3, 55
[Han17] Awni Hannun. Sequence modeling with ctc. Distill, 2(11):e8, 2017. 21
[Her82] H Herbert. The history of ocr, optical character recognition. Manchester Center,
VT: Recognition Technologies Users Association, 1982. 1
[HKR15] Samer Hijazi, Rishi Kumar, and Chris Rowen. Using convolutional neural net-
works for image recognition. Cadence Design Systems Inc.: San Jose, CA,
USA, pages 1–12, 2015. 81
[HM16] Birhanu Hailu and Million Meshesha. Applying image processing for malt-
barley seed identification. In Conference: Ethiopian the 9th ICT Annual Con-
ference, 2016. 18
[HML+ 18] Shizhong Han, Zibo Meng, Zhiyuan Li, James O’Reilly, Jie Cai, Xiaofeng
Wang, and Yan Tong. Optimizing filter size in convolutional neural networks
Bibliography 131
for facial action unit recognition. In Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition, pages 5070–5078, 2018. 19
[HOT06] Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algo-
rithm for deep belief nets. Neural computation, 18(7):1527–1554, 2006. 18
[HS95] Ke Han and Ishwar K Sethi. Off-line cursive handwriting segmentation. In

Proceedings of 3rd International Conference on Document Analysis and Recog-
nition, volume 2, pages 894–897. IEEE, 1995. 14
[HS06] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality

of data with neural networks. science, 313(5786):504–507, 2006. 18
[HZMJ15] Meijun He, Shuye Zhang, Huiyun Mao, and Lianwen Jin. Recognition confi-
dence analysis of handwritten chinese character with cnn. In 2015 13th Inter-
61–65. IEEE, 2015. 55
[HZRS16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learn-
ing for image recognition. In Proceedings of the IEEE conference on computer
vision and pattern recognition, pages 770–778, 2016. 18, 72
[IFD+ 19] R. R. Ingle, Y. Fujii, T. Deselaers, J. Baccash, and A. C. Popat. A scalable hand-
written text recognition system. In 2019 International Conference on Document
Analysis and Recognition (ICDAR), pages 17–24. IEEE, Sep. 2019. 3
[IKN98] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual
attention for rapid scene analysis. IEEE Transactions on pattern analysis and
machine intelligence, 20(11):1254–1259, 1998. 113
[Jam07] William James. The principles of psychology, volume 1. Cosimo, Inc., 2007. 22
[KJEY20] Mohamed Ibn Khedher, Houda Jmila, and Mounim A El-Yacoubi. Automatic
processing of historical arabic documents: A comprehensive survey. Pattern
Recognition, 100:107144, 2020. 2, 54
[KSBH16] Pavel Kisilev, Eli Sason, Ella Barkan, and Sharbell Hashoul. Medical image
captioning: learning to describe medical image findings using multi-task-loss
cnn. Deep Learning and Data Labeling for Medical Applications:in proceed-
ing of International Conference on Medical Image Computing and Computer-
Assisted Intervention, 2016. 63
[KSH12] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classifica-
tion with deep convolutional neural networks. In F. Pereira, C. J. C. Burges,
L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Pro-
cessing Systems 25, pages 1097–1105. Curran Associates, Inc., 2012. 18
[LBBH98] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-
based learning applied to document recognition. Proceedings of the IEEE,
86(11):2278–2324, 1998. 18, 43
132 Bibliography
[LFZ+ 15] Tianyi Liu, Shuangsang Fang, Yuehui Zhao, Peng Wang, and Jun Zhang.
Implementation of training convolutional neural networks. arXiv preprint
arXiv:1506.01195, 2015. 18
[LGF+ 07] Marcus Liwicki, Alex Graves, Santiago Fernàndez, Horst Bunke, and Jürgen
Schmidhuber. A novel approach to on-line handwriting recognition based on
bidirectional long short-term memory networks. In Proceedings of the 9th In-
ternational Conference on Document Analysis and Recognition, ICDAR 2007,
2007. 19, 20, 72
[LNNN17] Nam-Tuan Ly, Cuong-Tuan Nguyen, Kha-Cong Nguyen, and Masaki Naka-
gawa. Deep convolutional recurrent network for segmentation-free offline hand-
written japanese text recognition. In 2017 14th IAPR International Conference
on Document Analysis and Recognition (ICDAR), volume 7, pages 5–9. IEEE,
2017. 72, 91, 98
[LO16] Chen-Yu Lee and Simon Osindero. Recursive recurrent nets with attention mod-
eling for ocr in the wild. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 2231–2239, 2016. 98
[LPM15] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective

approaches to attention-based neural machine translation. arXiv preprint
arXiv:1508.04025, 2015. 19, 23, 24, 90, 92, 113, 117, 118, 119, 120
[LTD+ 17] Linghui Li, Sheng Tang, Lixi Deng, Yongdong Zhang, and Qi Tian. Image
caption with global-local attention. In Proceedings of the thirty-first AAAI con-
ference on artificial intelligence, pages 4133–4139, 2017. 27
[LV12] Hong Lee and Brijesh Verma. Binary segmentation algorithm for english cur-
sive handwriting recognition. Pattern Recognition, 45(4):1306–1317, 2012. 14
[MB02] U-V Marti and Horst Bunke. The iam-database: an english sentence database
for offline handwriting recognition. International Journal on Document Analy-
sis and Recognition, 5(1):39–46, 2002. 42
[MBP15] Durjoy Sen Maitra, Ujjwal Bhattacharya, and Swapan K Parui. Cnn based com-
mon approach to handwritten character recognition of multiple scripts. In Doc-
ument Analysis and Recognition (ICDAR), 2015 13th International Conference
on, pages 1021–1025. IEEE, 2015. 62
[Mes08] Million Meshesha. Recognition and retrieval from document image collections.
Ph. D. Dissertation, International Institute of Information Technology (IIIT),
2008. 3, 5, 14, 29, 31, 37, 54, 62, 63
[Mey06] Ronny Meyer. Amharic as lingua franca in ethiopia. Lissan: Journal of African
Languages and Linguistics, 20(1/2):117–132, 2006. 3, 30
[MF06] R Manmatha and Shaolei Feng. A hierarchical, hmm-based automatic evalu-

ation of ocr accuracy for a digital library of books. In Proceedings of the 6th
Bibliography 133
ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL’06), pages 109–

118. IEEE, 2006. 18
[Mil00] M Million. A generalized approach to optical character recognition of amharic
texts. Master’s thesis, School of Information Science, Addis Ababa University,
Addis Ababa, 2000. 38
[MJ05] Million Meshesha and CV Jawahar. Recognition of printed amharic documents.
In Eighth International Conference on Document Analysis and Recognition (IC-
DAR’05), pages 784–788. IEEE, 2005. 29
[MJ07] Million Meshesha and CV Jawahar. Optical character recognition of amharic
documents. African Journal of Information & Communication Technology,
3(2), 2007. 14, 30, 32, 54, 55, 56, 60, 61, 62, 72, 73, 98
[ML15] Ronaldo Messina and Jerome Louradour. Segmentation-free handwritten chi-
nese text recognition with lstm-rnn. In 2015 13th International Conference on
Document Analysis and Recognition (ICDAR), pages 171–175. IEEE, 2015. 72,
98
[MM18] Getahun Tadesse Mekuria and Getahun Tadesse Mekuria. Amharic text docu-
ment summarization using parser. International Journal of Pure and Applied
Mathematics, 118(24), 2018. 4, 30
[MSLB98] John Makhoul, Richard Schwartz, Christopher Lapre, and Issam Bazzi. A
script-independent methodology for optical character recognition. Pattern
Recognition, 31(9):1285–1294, 1998. 18
[MZ17] Tirin Moore and Marc Zirnsak. Neural mechanisms of selective visual attention.
Annual review of psychology, 68:47–72, 2017. 113
[Nag92] George Nagy. At the frontiers of ocr. Proceedings of the IEEE, 80(7):1093–
1100, 1992. 17
[Nag00] George Nagy. Twenty years of document image analysis in pami. IEEE Trans-
actions on Pattern Analysis & Machine Intelligence, (1):38–62, 2000. 14
[NAL] The ethiopian national archive and liberary agency. http://www.nala.
gov.et/home. Accessed on 10 Feburary 2021. 5
[NWC+ 11] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and An-
drew Y Ng. Reading digits in natural images with unsupervised feature learn-
ing. In NIPS workshop on deep learning and unsupervised feature learning,
page 5, 2011. 43
[Pal05] Mahesh Pal. Random forest classifier for remote sensing classification. Inter-
national journal of remote sensing, 26(1):217–222, 2005. 18
[PC00] U Pal and BB Chaudhuri. Automatic recognition of unconstrained off-line
bangla handwritten numerals. In International Conference on Multimodal In-
terfaces, pages 371–378. Springer, 2000. 31
134 Bibliography
[PC04] U Pal and BB Chaudhuri. Indian script character recognition: a survey. pattern
Recognition, 37(9):1887–1899, 2004. 14
[PPCA13] Christos Papadopoulos, Stefan Pletschacher, Christian Clausner, and Aposto-

los Antonacopoulos. The impact dataset of historical document images. In
Proceedings of the 2nd international workshop on historical document imaging
and processing, pages 123–130, 2013. 43
[pre] The ethiopian press agency. will amharic be au’s lingua franca ? https://
www.press.et/english/?p=2654#l. Accessed on 10 November 2019.
30
[PV17] Jason Poulos and Rafael Valle. Character-based handwritten text transcription
with attention networks. arXiv preprint arXiv:1712.04046, 2017. 27, 98
[PZLL20] Xushan Peng, Xiaoming Zhang, Yongping Li, and Bangquan Liu. Research
on image feature extraction and retrieval algorithms based on convolutional
neural network. Journal of Visual Communication and Image Representation,
69:102705, 2020. 19
[RBO17] Christophe Rigaud, Jean-Christophe Burie, and Jean-Marc Ogier.

Segmentation-free speech text recognition for comic books. In 2017
th
14 IAPR International Conference on Document Analysis and Recognition
(ICDAR), volume 3, pages 29–34. IEEE, 2017. 72
[RMS09] Amjad Rehman, Dzulkifli Mohamad, and Ghazali Sulong. Implicit vs ex-
plicit based script segmentation and recognition: a performance comparison
on benchmark database. Int. J. Open Problems Compt. Math, 2(3):352–364,
2009. 15
[RRB18] Betselot Yewulu Reta, Dhara Rana, and Gayatri Viral Bhalerao. Amharic hand-
written character recognition using combined features and support vector ma-
chine. In 2018 2nd International Conference on Trends in Electronics and In-
formatics (ICOEI), pages 265–270. IEEE, 2018. 38, 62, 63, 72, 73
[RSWP18] Christian Reul, Uwe Springmann, Christoph Wick, and Frank Puppe. State of
the art optical character recognition of 19th century fraktur scripts using open
source engines. arXiv preprint arXiv:1810.03436, 2018. 73
[RTAS20] António H Ribeiro, Koen Tiels, Luis A Aguirre, and Thomas Schön. Beyond
exploding and vanishing gradients: analysing rnn training using attractors and
smoothness. In International Conference on Artificial Intelligence and Statis-
tics, pages 2370–2380, 2020. 19
[SA15] Maad Shatnawi and Sherief Abdallah. Improving handwritten arabic character
recognition by modeling human handwriting distortions. ACM Transactions on
Asian and Low-Resource Language Information Processing (TALLIP), 15(1):1–
12, 2015. 56
Bibliography 135
[SG97] Zhixin Shi and Venu Govindaraju. Segmentation and recognition of connected
handwritten numeral strings. Pattern Recognition, 30(9):1501–1504, 1997. 14
[SIK+ 09] Fouad Slimane, Rolf Ingold, Slim Kanoun, Adel M Alimi, and Jean Hennebert.
A new arabic printed text image database and evaluation protocols. In 2009
10th International Conference on Document Analysis and Recognition, pages
946–950. IEEE, 2009. 42
[SJS13] P Sibi, S Allwyn Jones, and P Siddarth. Analysis of different activation func-
tions using back propagation neural networks. Journal of theoretical and ap-
plied information technology, 47(3):1264–1268, 2013. 18
[SLJ+ 15] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,
Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabi-
novich. Going deeper with convolutions. In Proceedings of the IEEE confer-
ence on computer vision and pattern recognition, pages 1–9, 2015. 18
[SS13] Nazly Sabbour and Faisal Shafait. A segmentation-free approach to arabic
and urdu ocr. In Document Recognition and Retrieval XX, volume 8658, page
86580N. International Society for Optics and Photonics, 2013. 42
[SSSS17] Shai Shalev-Shwartz, Ohad Shamir, and Shaked Shammah. Failures of
gradient-based deep learning. In Proceedings of the 34th International Con-
ference on Machine Learning-Volume 70, pages 3067–3075. JMLR. org, 2017.
86
[STA+ 18] Palaiahnakote Shivakumara, Dongqi Tang, Maryam Asadzadehkaljahi, Tong

Lu, Umapada Pal, and Mohammad Hossein Anisi. Cnn-rnn based method
for license plate recognition. CAAI Transactions on Intelligence Technology,
3(3):169–175, 2018. 91
[SUHP+ 15] Fotini Simistira, Adnan Ul-Hassan, Vassilis Papavassiliou, Basilis Gatos, Vas-
silis Katsouros, and Marcus Liwicki. Recognition of historical greek polytonic
scripts using lstm networks. In 2015 13th International Conference on Docu-
ment Analysis and Recognition (ICDAR), pages 766–770. IEEE, 2015. 73
[SVL14] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning
with neural networks. In Advances in neural information processing systems,
pages 3104–3112, 2014. 24, 113, 115
[SZ14] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for
large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 18, 55
[Tef99] Dereje Teferi. Optical character recognition of typewritten amharic text. Mas-
ter’s thesis, School of Information Science, Addis Ababa University, Addis
Ababa, 1999. 38, 54, 55, 60, 62, 63, 72, 73
[Tes10] Abay Teshager. Amharic character recognition system for printed real-life doc-
uments. Master’s thesis, School of Information Science, Addis Ababa Univer-
sity, Addis Ababa, 2010. 32
136 Bibliography
[TT18] Zhiqiang Tong and Gouhei Tanaka. Reservoir computing with untrained con-
volutional neural networks for image recognition. In 2018 24th International
Conference on Pattern Recognition (ICPR), pages 1289–1294. IEEE, 2018. 19
[UH16] Adnan Ul-Hasan. Generic text recognition using long short-term memory net-
works. Ph. D. Dissertation, Technical University of Kaiserslautern, 2016. 17
[UHAR+ 13] Adnan Ul-Hasan, Saad Bin Ahmed, Faisal Rashid, Faisal Shafait, and
Thomas M Breuel. Offline printed urdu nastaleeq script recognition with bidi-
rectional lstm networks. In 2013 12th International Conference on Document
Analysis and Recognition, pages 1061–1065. IEEE, 2013. 73
[VMN+ 16] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie.
Coco-text: Dataset and benchmark for text detection and recognition in natural
images. In arXiv preprint arXiv:1601.07140, 2016. 43
[Wer90] Paul J Werbos. Backpropagation through time: what it does and how to do it.
Proceedings of the IEEE, 78(10):1550–1560, 1990. 19
[WHK+ 17] Shinji Watanabe, Takaaki Hori, Suyoun Kim, John R Hershey, and Tomoki
Hayashi. Hybrid ctc/attention architecture for end-to-end speech recognition.
IEEE Journal of Selected Topics in Signal Processing, 11(8):1240–1253, 2017.
90
[Wio06] Anais Wion. The national archives and library of ethiopia: six years of ethio-
french cooperation (2001-2006). 2006. 4
[WSW+ 17] Qi Wu, Chunhua Shen, Peng Wang, Anthony Dick, and Anton van den Hen-
gel. Image captioning and visual question answering based on attributes and
external knowledge. IEEE transactions on pattern analysis and machine intel-
ligence, 40(6):1367–1381, 2017. 20
[WYCL17] Yi-Chao Wu, Fei Yin, Zhuo Chen, and Cheng-Lin Liu. Handwritten chinese
text recognition using separable multi-dimensional recurrent neural network. In
2017 14th IAPR International Conference on Document Analysis and Recogni-
tion (ICDAR), volume 1, pages 79–84. IEEE, 2017. 72, 98
[XBK+ 15] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan
Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural
image caption generation with visual attention. In International conference on
machine learning, pages 2048–2057, 2015. 23, 24, 117, 118, 119
[YBJL12] Aiquan Yuan, Gang Bai, Lijing Jiao, and Yajie Liu. Offline handwritten en-
glish character recognition based on convolutional neural network. In 2012 10th
IAPR International Workshop on Document Analysis Systems, pages 125–129.
IEEE, 2012. 54, 55, 91
[YNDT18] Rikiya Yamashita, Mizuho Nishio, Richard Kinh Gian Do, and Kaori Togashi.
Convolutional neural networks: an overview and application in radiology. In-
sights into imaging, 9(4):611–629, 2018. 19
Bibliography 137
[YZX+ 18] Min Yang, Wei Zhao, Wei Xu, Yabing Feng, Zhou Zhao, Xiaojun Chen, and Kai
Lei. Multitask learning for cross-domain image captioning. IEEE Transactions
on Multimedia, 2018. 63
[ZCW10] Shusen Zhou, Qingcai Chen, and Xiaolong Wang. Hit-or3c: an opening recog-
nition corpus for chinese characters. In Proceedings of the 9th IAPR Interna-
tional Workshop on Document Analysis Systems, pages 223–230, 2010. 43
[ZDD18] Jianshu Zhang, Jun Du, and Lirong Dai. Track, attend, and parse (tap): An end-
to-end framework for online handwritten mathematical expression recognition.
IEEE Transactions on Multimedia, 21(1):221–233, 2018. 20, 27, 98
[ZHZ17] Haifeng Zhao, Yong Hu, and Jinxia Zhang. Character recognition via a compact
convolutional neural network. In 2017 International conference on digital im-
age computing: techniques and applications (DICTA), pages 1–6. IEEE, 2017.
55, 56
[ZLLT14] Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaoou Tang. Facial land-
mark detection by deep multi-task learning. In European Conference on Com-
puter Vision, pages 94–108. Springer, 2014. 63
[ZY17] Yu Zhang and Qiang Yang. A survey on multi-task learning. arXiv preprint
arXiv:1707.08114, 2017. 64
[ZYDS17] Geoffrey Zweig, Chengzhu Yu, Jasha Droppo, and Andreas Stolcke. Advances
in all-neural speech recognition. In 2017 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), pages 4805–4809. IEEE,
2017. 20
Curriculum Vitae
Full Name: Birhanu Hailu Belay
Work Experience
October 2017–Now PhD Student TU Kaiserslautern, DFKI, Kaiserslautern, Germany
January 2013 Assistant Lecture, Bahir Dar Institute of Technology, Ethiopia

– September 2017
September 2011 Graduate Assistant, Deberetabor University, Ethiopia
– December 2012
September 2010 Graduate Assistant, Arbaminch University, Ethiopia
– August 2011
Education
October 2017 PhD in Computer Science, TU Kaiserslautern, Germany
– Now Thesis: ”Deep Learning for Amharic Text-image Recognition: Algorithm, Dataset
and application”
September 2013 Master of Science in Computer Science,

– October 2015 Bahir Dar Institute of Technology, Ethiopia
Thesis: ”Applying Image processing techniques for Malt-barley Seed Identification”
September 2007 Bachelor of Science Information Communication Technology,

– July 2010 Wollo University, Ethiopia
Thesis/project: ”Patient Registration and Medical Records Management System”
Computer Skill
Programming Python, Java, C++

Web development CSS, HTML, PHP, Javascript, JSP
Database Microsoft SQL Server, Microsoft Access
Development Environments Matlab, Macromedia Dreamweaver 8, VisualStudio
Operating Systems Linux and Windows operating system
Documentation softwares MS Office products, Latex
Deep learning frameworks Keras, Tensorflow
Publication list
Journal articles
1. Birhanu Hailu Belay, Tewodros Habtegebrial, Million Meshesha, Marcus

Liwicki, Gebeyehu Belay, Didier Stricker, ”Amharic OCR: An End-to-End
Learning”. MDPI-Applied Sciences. 2020; 10(3):1117.
2. Birhanu Hailu Belay, Tewodros Habtegebrial, Gebeyehu Belay, Million Meshe-

sha, Marcus Liwicki, Didier Stricker, ”Learning by Injection: Attention Em-
bedded Recurrent Neural Network for Amharic Text-image Recognition”. Preprints,
2020
Conference papers

Belay, Didier Stricker, ”Blended Attention-CTC for Amharic Text-image Recog-
nition: Sequence-to-Sequence Learning”, 10th International Conference on
Pattern Recognition Application and Methods (ICPRAM 2021)
4. Birhanu Hailu Belay, Tewodros Habtegebrial, Gebeyehu Belay, Didier Stricker,

”Using Automatic Features for Text-image Classification in Amharic Docu-
ments”, 9th International Conference on Pattern Recognition Applications and
Methods. International Conference on Pattern Recognition Applications and
Methods (ICPRAM-2020) February 22-24 Valletta, Malta 2020.

Belay, Didier Stricker, ”Factored Convolutional Neural Network for Amharic
Character Image Recognition”, 26th IEEE International Conference on Image
Processing (ICIP), Taipei, Taiwan, 22–25 September 2019; pp. 2906–2910.

Belay, Didier Stricker, ”Amharic Text Image Recognition: Dataset, Algorithm
and Analysis”, 15th International Conference Document Analysis and Recog-
nition (ICDAR), Sydney, Australia, 20–25 September 2019; pp. 1268–1273.
7. Birhanu Hailu Belay, Tewodros Habtegebrial, Gebeyehu Belay, Didier Stricker,

”Multi-font Printed Amharic Character Image Recognition: Deep Learning
Techniques”, 6th EAI International Conference on Advances of Science and
Technology: ICAST 2018, Bahir Dar, Ethiopia, October 5-7, 2018, Proceed-
ings, vol. 274, p. 322. Springer, 2019
8. Birhanu Hailu Belay, Tewodros Habtegebrial, Didier Stricker, ”Amharic char-
acter image recognition”, 18th IEEE International Conference on Communica-
tion Technology (ICCT), Chongqing, China, 8–11 October 2018; pp. 1179–1182
9. Birhanu Hailu Belay, Million Meshesha: “Application of Image Processing
Techniques for Malt-barley Seed Identification”, 9th ICT Annual Conference
2016 (EICTAC 2016)
10. Birhanu Hailu Belay, Tesfa Tegegne, ”Ethiopian Roasted Coffee Classifi-
cation Using Imaging Techniques”, 3rd International Conference on the Ad-
vancement of Science and Technology (ICAST2015), pp. 23-29, Bahir Dar,
Ethiopia, 2015.

Deep Learning For Amharic Text-ImageRecognition

Uploaded by

Copyright:

Available Formats

Deep Learning For Amharic Text-ImageRecognition

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deep Learning For Amharic Text-ImageRecognition

Uploaded by

Copyright:

Available Formats

Deep Learning for Amharic Text-Image

Recognition: Algorithm, Dataset and Application

Birhanu Hailu Belay

Date of Defense: 10.02.2021

List of figures xii

List of tables xiii

2 Trends in OCR Research and Development 13

3 Amharic Script and Writing System 29

3.4 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4 Amharic OCR and Baseline Datasets 41

5 Recognition of Amharic Characters 53

6 Amharic Script Recognition using CTC Sequence Modeling 71

6.2.5 Recognition Performance Comparison . . . . . . . . . . . . . . 86

7 Attention-based Sequence Learning for Amharic OCR 89

8 Conclusions and Recommendations for Future Work 105

1.1 A generic Document Image Analysis (DIA) flow diagram . . . . . . . 2

2.1 Character segmentation process and segmentation problems on sample

3.1 Demographic distribution of Amharic speakers . . . . . . . . . . . . . 30

4.1 Sample synthetically generated Amharic character images . . . . . . . 44

5.1 A CNN-based OCR model for Amharic character image recognition . . 58

6.1 The proposed network architecture . . . . . . . . . . . . . . . . . . . . 75

6.3 Training versus validation loss of an LSTM-CTC model . . . . . . . . 77

7.1 Attention-based Encoder-Decoder model for Amharic OCR . . . . . . . 94

A.1 General setting of Attention model . . . . . . . . . . . . . . . . . . . . 114

5.1 Character level recognition accuracy (out of 100%) and performance

6.1 The detail of Amharic text-line images database . . . . . . . . . . . . . 74

OCR Optical Character Recognition

The history of Optical Character Recognition (OCR) is traced back to the

Optical Character Recognition (OCR) refers to a set of computer vision

the amount of paper-based documents. A large number of documents, hav-

• Feature extraction: The distinctive characteristics of the image, useful

• Post-processing: At this step, the recognition errors are corrected based

related to religion, history, literature, politics, economics, philosophy, aca-

Nowadays, document processing and preservation has been given much

1.2 Statement of the Problem

state-of-the-art deep learning algorithms. In parallel, we develop a baseline

1.3 Research Objectives

1. The performance of OCR model for basic Amharic character recogni-

2. The ability of RNNs to process contextual information enables them

3. In NMT, Attention enhances the ability of the networks in extracting the

• A well structured and standard database of basic Amharic character im-

• Created the first publicly available Amharic text-line image dataset.

• Propose a novel Approach for basic Amharic character recognition. To

• Designed an OCR for recognizing Amharic scripts at text-line level.

• Adoption of synthetic database generation methodology and propose

• The basic idea of Attention mechanism, on neural machine translation,

• Propose a novel sequence-to-sequence learning method by blending the

• A detail analysis of the results of experiments carried out with different

1.5 Organization of the Thesis

Chapter three introduces Amharic languages and presents background in-

Chapter five discusses CNN-based Amharic character recognition models

Finally, a summary of this thesis, concluding remarks, and guidelines for

1.6 List of Publications

1. Birhanu Hailu Belay, Tewodros Habtegebrial, Million Meshesha, Mar-

Conferences and Preprints

1. Birhanu Hailu Belay, Tewodros Habtegebrial, Marcus Liwicki, Gebeyehu

Amharic Character Image Recognition”, 26th IEEE International Con-

2.1 Classical OCR Approach

The classical OCR approach is commonly called the segmentation-based ap-