HANDWRITTEN GUJARATI CHARACTER RECOGNITION BASED

ON DISCRETE COSINE TRANSFORM

1
MADHUREE ARDESHANA, 2ANKIT K. SHARMA, 3DIPAK M. ADHYARU, 4TANISH H. ZAVERI
1,2,3,4
Institute of Technology, Nirma University, Ahmedabad, Gujarat, India
E-mail: ardeshana.g27@gmail.com, 2ankit.sharma@nirmauni.ac.in, 3dipak.adhyaru@nirmauni.ac.in,
1
4
ztanish@nirmauni.ac.in

Abstract— Gujarati script is used by more than 600 million people around the world. Handwritten character recognition of
Gujarati script is important area of research due to its wider applications. The work done for handwritten Gujarati character
recognition is negligible in literature. Here we have suggest a method based on Discrete Cosine Transform (DCT) as feature
extraction method and Naïve Bayes (NB) classifier for classification of characters. For experiment database used here is
around 22000 samples. The result obtained from above setup gives good accuracy.

Keywords—Handwritten Gujarati characters, Discrete cosine transform, Naïve Bayes classifier

I. INTRODUCTION used here. We believe that the work presented here

will be useful for the further development of OCR
Optical Character Recognition (OCR) system can system of Gujarati script having high accuracy.
convert scanned document into text format for
storage, editing, transmission, indexing and
coordinating into different applications.
Advancement of OCR innovations of Indian content
has more challenges than western content due to the
perplexing character set and also presence of joint
characters and modifiers [1]. As Gujarat is one of the
eminent state of India, Gujarati is a well-known and
culturally rich dialect. Therefore improvement in
effective OCR of Gujarati dialect will contribute
more. Gujarati script is obtained from antique
Devanagari script and it has much likeness with other
north Indian dialect, basically Hindi. The significant
distinction between other north Indian dialect and
Gujarati is the nonappearance of a head line going
through every characters framing the word [1]. Fig. 1 Set of handwritten Gujarati characters
Because of its unconventional characteristics Gujarati
should to be treated by different approach from other This paper describes all important stages of OCR
Indo-Aryan dialects. Gujarati Character Recognitions system. Figure 2 shows the general architecture of
offers more difficulties like most other Indian scripts OCR model. For any OCR system, feature extraction
relative to the western languages due to these reasons: and classification stages are important. The rest of
(a) presence of joint characters, (b) presence of the paper is organized as below. Section II and
similar looking characters, (c) huge and complex Section III discuss about preprocessing and
character set by combination of characters and segmentation stage respectively. Dataset preparation
modifiers [2]. Set of Gujarati characters in and information about it is discussed in section IV.
handwritten form is shown in fig. 1.1 which includes Feature extraction and experiment result are
basic consonants and vowels of Gujarati language. discussed in section V and VI respectively.
For Gujarati character recognition very less
significant work is available in the literature. These
are also restricted to finite number of images for
recognition [3]. For researchers Handwritten
character recognition has turned into an extremely
fascinating topic since most recent couple of decades
as it is very challenging area due to variations in
writing style.We have performed the experiments for
large set of character images and provided a
significant outcome for handwritten Gujarati
character recognition. This paper depicts the
outcomes acquired by using the Naïve Bayes
classifier. The Discrete Cosine Transform features are Fig. 2 OCR architecture

Proceedings of IRF-ieeeforum International Conference, 3rd April, 2016, Pune, India, ISBN: 978-93-85973-95-6
23
 Handwritten Gujarati Character Recognition Based on Discrete Cosine Transform

II. PREPROCESSING Slant angle between a vertical direction and longest

stroke in a word is one of the quantifiable factor of
The handwritten characters in crude form will be various handwriting styles. Slant removal is used for
subjected to various preprocessing stages to make it normalization of each characters in a standard form
usable regardless how it is gained. The preprocessing [11].
stage aims to produce the data relevant for the
recognition stage. The principle goals of pre- III. SEGMENTATION
processing are:
Segmentation is one of the most imperative part of
(1) Binarization:- preprocessing step. It permits the classifier to extract
For any grayscale captured image every pixel varies information from every characters. For handwritten
between 0 to 255 value. With binarization process character recognition this process is very crucial as
greyscale value is thresholded and it is being characters have a tendency to be joined with one
converted into black(‘1’)for foreground or white(‘0’) another, distorted or overlapped. For complex and
for background pixel. The method used for joint strings advanced methods must be utilized.
binarization of any image is called thresholding in Segmentation is carried out for breaking a single text
which gray scale image is converted into binary by a line from scanned documents, single word from
finite threshold value. Thresholding is divided into single lines, and single character from the single
two groups which are local and global thresholding. word. Segmentation can be classified majorly in two
Adaptive or local thresholding utilizes different classes [11]:
threshold values for every pixel based on local area a) External segmentation, which is performed for
information. Whereas in global thresholding , for isolation of paragraphs, single lines or words.
entire image one threshold value is utilized which is b) Internal segmentation, which is performed for
estimated by intensity histogram. Local thresholding isolation of single characters.
is generally useful for images having different level
of intensities, for example, images from satellite There are many techniques available for segmentation
cameras. Whreaeas global thresholding would be of individual characters which are based on
sufficient for simple images such as handwriting on projection profiles, connected component labeling or
white background [10]. white space and pitch [12].

(2) Noise Reduction:- IV. DATASET

Any scanned document may have noises for example,
holes in lines, disconnected segments etc. for Remarkable research works are available on printed
improving quality of the document noise reduction Gujarati script. However, a very few work is
process is performed.[11] There are more than available on handwritten Gujarati script. For research
hundreds of techniques available for noise reduction work on handwritten Gujarati characters needs
these are mainly classified into three groups: filtering, proper standard or benchmark databases. So our
noise modeling and morphological operations. Filters research work also incorporates to develop a large
can be intended for sharpening, smoothing, contrast anddelegate sample databases for handwritten
adjustment etc. Different morphological operations Gujarati script. Database was gathered by filling an
can be intended for thinning the characters, extracting application form by distinctive gatherings of the
boundaries, smoothing the contour etc [10]. students of different age. Application forms were
gathered and isolated handwritten characters were
(3) Normalization:- extracted from the details mentioned in application
Through normalization process images with random forms by students. All these application forms were
size is converted into standard size. Here, scanned through HP Scanjet 3600 multipage scanner
normalization in size maintain inter class invariation at 300 dpi resolutions. For handwritten Gujarati
of characters. A couple of methods are available for characters more than 22000 character samples were
size normalization for example, bi-cubic extracted from these forms. These individual
interpolation, bilinear etc [12]. Gujarati character samples were saved in JPG format.
Extraction of isolated characters from the scanned
(4) Skew Correction:- documents manually takes lot of time so for saving
This process is carried out for aligning the scanned time by software isolation of characters is done.
document with coordinate system of the scanner. Segmentation of line and word and characters are
Generally correlation, projection profiles etc methods done. Thus these database in not generated in
are used for skew detection [11]. laboratory environments. The database is equally
distributed among different classes for correlation
(5) Slant Removal:- purposes. Our database comprises of 22000 samples
Writing style varies from person to person therefore and total 44 classes. This Database of each class is
the slant of handwritten characters will be different. also divided in training part and testing part in ratio of

Proceedings of IRF-ieeeforum International Conference, 3rd April, 2016, Pune, India, ISBN: 978-93-85973-95-6
24
 Handwritten Gujarati Character Recognition Based on Discrete Cosine Transform

5:1. So The entire arrangement of available data of CONCLUSIONS

each individual class have been part into a training set
a test set. The samples are stored in grayscale In this paper we have implemented a NB classifier for
images. So that the researchers can explore different Gujarati Handwritten character recognition system.
techniques of preprocessing, thresholding etc. The features used for recognition is DCT which is
easy to obtain. Here we have obtained good
V. FEATURE EXTRACTION recognition rate of 78.05 % for large dataset.We hope
that this work will also be useful for the research
For increasing the recognition rate of classifier, work of optical character recognition for other Indian
unique features are computed from each individual scripts.
characters. Feature extraction method transforms the
input data into the set of features called feature vector ACKNOWLEDGMENTS
which is a reduced representation of the input data
[8]. Feature extraction is very crucial stage. Features Authors acknowledge the Institute of Technology,
are extracted after preprocessing and segmentation Nirma University for providing the research facilities.
stage. Here we have used Discrete Cosine Transform
for extracting features from input data. The DCT is a REFERENCES
linear transformation which expresses a sequence of
finitely many data points which is a sum of cosine [1] MukeshGoswami and SumanMitra, P. Maji et al,”
functions that oscillates at different frequencies which Structural Feature Based Classification of Printed Gujarati
Characters ” , In Springer-Verlag Berlin Heidelberg,
preserve the most useful features. DCT transforms an LNCS 8251, 2013.
n-dimensional vector into the set of n coefficients. [2] MandarChaudhary, GitamShikkenawis, Suman K. Mitra,
This transformation is spatial domain to the MukeshGoswami, “Similar looking Gujarati printed
character recognition using Locality Preserving Projection
frequency domain. It has only real transform domain
and Artificial Neural Networks” ,In third International
coefficients and includes only positive frequencies In Conference on Emerging Applications of Information
an image, the lower frequencies stores most of energy Technology (EAIT), IEEE , 978-1-4673-1827-3/12 2012.
so by transforming an image into frequency domain, [3] JigneshDholakia, ArchitYajnik, AtulNegi,” Wavelet
Feature Based Confusion Character Sets for Gujarati
we are able to reduction in data that describe the
Script”,In International Conference on Computational
image with sufficient quality.For an image, two- Intelligence and Multimedia Applications, IEEE ,0-7695-
dimensional(2-D) DCT need to be used as the input 3050-8/07, 2007.
data are two- dimensional. The 2-D DCT can be [4] Desai AA , “Gujarati handwritten numeral optical
derived from one-dimensional (1-D) DCT because character recognition through neural network”. Pattern
Recognition, 2582–2589, 2010.
each dimension of an image can be handled [5] Patel CN, Desai AA , “Segmentation of text lines into
separately. For that A one-dimensional DCT is words for Gujarati handwritten text. In: Proceedings of
computed by the length of the columns and then international conference on signal and image processing”,
In IEEE Xplore,2010
along with the rows or vice versa which means
[6] Patel CN, Desai AA, “Zone identification for Gujarati
operation is applied 1D DCT horizontally to rows and handwritten words. In: Proceedings of international
then apply 1D DCT vertically to resultant horizontal conference on emerging applications of information
DCT.With DCT, each character image is represented technology”, In IEEE Xplore, EAIT 2011
[7] Patel CN, Desai AA, “Gujarati handwritten character
as one vector. DCT can also convert the energy of an recognition using hybrid method based on binary tree-
image into a few coefficients. By applying DCT on classifier and k-nearest neighbour” , In J Eng Res Technol
the character image with size 16x16, 256 DCT , 2013.
coefficients of the image are obtained. The number of [8] Lipi Shah, Ripal Patel, Shreyal Patel, Jay Maniar,”
Extensive experiment was performed on the whole Handwritten Character Recognition using
RadialHistogram”,InInternational Journal of Research in
database of 22,000 character images. Advent Technology, E-ISSN: 2321-9637,2014.
[9] Hetal R. Thaker, C. K. Kumbharana,” Structural Feature
VI. EXPERIMENT RESULTS Extraction to recognize some of the Offline Isolated
Handwritten Gujarati Characters using Decision Tree
Classifier”,In International Journal of Computer
The features obtained here is DCT and it is utilized Applications, 2014.
by the classifier. Naïve Bayes classifier is used here [10] AzizahSuliman, Mohd. NasirSulaiman, Mohamed
for classification. The database used for Othman, RahmitaWirza ,” Chain Coding and Pre
Processing Stages of Handwritten Character Image File”.
experimentation is too large from any other [11] Gaurav Y. Tawde , Mrs. Jayashree M. Kundargi , “An
approaches done before for the handwritten Gujarati Overview of Feature Extraction Techniques in OCR for
character recognition.There are 500 images for each Indian Scripts Focused on Offline Handwriting”.
characters and experiment is performed on 22000 [12] Dr. Jangala. SasiKiran, N. VijayaKumar , N. SashiPrabha,
samples. Overall accuracy for proposed scheme is M. Kavya,” A Literature Survey on Digital Image
Processing Techniques in Character Recognition of Indian
found 78.05% from Naïve Bayes classifier. From the Languages”.
result of the NB classifier, we can conclude that even [13] IsraaHadiAli , “New Method for Image Features
though classifier used here is simple it delivers good Extracting Based on Enhanced Chain Code-”,In
ORIENTAL JOURNAL OF COMPUTER SCIENCE &
recognition rate with DCT features.
TECHNOLOGY,2013.

Proceedings of IRF-ieeeforum International Conference, 3rd April, 2016, Pune, India, ISBN: 978-93-85973-95-6
25
 Handwritten Gujarati Character Recognition Based on Discrete Cosine Transform
[14] Satish Kumar, “Performance Comparison of Features on [17] Mr. Mukesh M. Goswami, Mr. Harshad B. Prajapati, Mr.
Devanagari Hand-printed Dataset”,In International Journal Vipul K. Dabhi,” Classification of Printed Gujarati
of Recent Trends in Engineering, Vol. 1, No. 2, May 2009. Characters using SOM based K-Nearest Neighbor
[15] Ian H. Witten and Eibe Frank, Data Mining, “Practical Classifier”,In International Conference on Image
Machine Learning Tools and Techniques”,In Morgan Information Processing (ICIIP) Proceedings of the 2011
Kaufmann Publication, 2005. International Conference on Image Information Processing
[16] Jiawei Han and MichelineKamber,” Data Mining: (ICIIP), 978-1-61284-861-7/11, 2011.
Concepts and Techniques”,In Morgan Kaufmann
Publication, 2001.



Proceedings of IRF-ieeeforum International Conference, 3rd April, 2016, Pune, India, ISBN: 978-93-85973-95-6
26