2

Abstract

The world in the 21st century is ever evolving towards automation. This upsurge seemingly has
no decline in the foreseeable future. Image recognition is at the forefront of this charge which
seeks to revolutionize the way of living of the average man. If robotics can be likened to the cre-
ation of a body for computers to live in, then image processing is the development of the part of
its brain which deal with identification and recognition of images.
To accomplish this task, we developed an object detection algorithm using YOLO, and acronym
for “You Only Look Once”. Our algorithm was trained on fifty thousand images and evaluated
on ten thousand images and employed a 21 x 21 grid. We also programmed a text generator
which randomly creates texts and URLs in an image. A record of useful information about the
location of the URLs in the image is also recorded and later passed to the YOLO algorithm for
training.
At the end of this project, we observed significant difference in the accuracy of URL detection
when using an OCR software or our YOLO algorithm. However, our algorithm would be best
used to specify the region of interest before converting to texts which greatly improves accuracy
when combined with OCR software.

Keywords: Object Detection; Image Recognition; OCR; AI; World Wide Web; YOLO.
 3

Acknowledgements

I would like to express my deepest appreciation to God for life. I would also like to ex-

tend my deepest gratitude to all the Professors at St. Cloud State University, who helped make

this academic journey worthwhile. Learning Software Development concepts, Database con-

cepts, Data Structure and Algorithm concepts, Machine Learning concepts, and other important

concepts relevant to “real-time” applications of Software Development was an awesome experi-

ence. I am extremely grateful for everything I got to learn in this University.

I would like to thank my friends, Alienyi Daniel, Ojo Ore, Egwu Benedict, Nwachi Eurel,

et al. for their constant support and encouragement. Finally, and most importantly, I want to

thank my family for believing in me and constantly encouraging me throughout this journey.
 4

Table of Contents

Page

List of Figures ............................................................................................................................. 6

Chapter

I. Introduction ..................................................................................................................... 8

Overview ................................................................................................................. 8

Motivation ............................................................................................................... 8

Research aim ........................................................................................................... 9

Problem description and Research Questions ......................................................... 9

What is an image in Computer Science? ................................................................ 9

What is Object Detection? .................................................................................... 13

What is Image Recognition? ................................................................................. 14

II. Background ................................................................................................................... 17

Overview ............................................................................................................... 17

Image Processing .................................................................................................. 17

Artificial intelligence and learning ....................................................................... 18

Machine Learning ................................................................................................. 19

Machine Learning Evaluation ............................................................................... 21

Deep Learning ....................................................................................................... 22

Neural Network ..................................................................................................... 23

Unified Detection Model – YOLO ....................................................................... 28

 5

Chapter Page

Pytesseract............................................................................................................. 39

Regular Expressions.............................................................................................. 41

III. Experiment .................................................................................................................... 43

Overview ............................................................................................................... 43

Data Preparation.................................................................................................... 43

Training ................................................................................................................. 45

Evaluation Metric.................................................................................................. 47

IV. Results ........................................................................................................................... 49

Overview ............................................................................................................... 49

Detection Accuracy ............................................................................................... 49

Font Style and Size ............................................................................................... 52

Input Size .............................................................................................................. 54

V. Conclusion..................................................................................................................... 55

Overview ............................................................................................................... 55

Application ............................................................................................................ 55

Limitations of Work .............................................................................................. 57

Future work ........................................................................................................... 57

References ................................................................................................................................. 59
 6

List of Figures

Figure Page

1. Image Recognition and Object Detection difference [36]. ............................................... 16

2. Enhancing grayscale images with histogram equalization [37]. ....................................... 17

3. Structure of Perceptron [35].............................................................................................. 23

4. Output of a Perceptron [35]. ............................................................................................. 24

5. Activation functions. (1) the left curve is a sigmoid function curve. (2) the right curve is a

tanh function curve [35]. ................................................................................................... 24

6. A sample fully connected neural network with only one hidden layer [35]. .................... 25

7. Visualization of a Training Process [35]. ......................................................................... 26

8. Max-pooling. Pooling from 24 × 24 to 12 × 12 [35]. ....................................................... 27

9. Structure of a CNN Based Object Detection Model [35]. ................................................ 28

10. YOLO Structure [35]. ....................................................................................................... 31

11. YOLO Network Architecture [42]. ................................................................................... 31

12. Bounding Box prediction formula [32]............................................................................. 36

13. Image before pre-processing [43]. .................................................................................... 40

14. Image after pre-processing [43]. ....................................................................................... 41

15. Word bank used to generate our training image data. ...................................................... 44

16. Sample generated image using our custom-built tagging tool. ......................................... 44

17. Sample grid generated by YOLO for our detection algorithm. ........................................ 45

18. Our detection model summary. ......................................................................................... 47

 7

Figure Page

19. Graphical View of the IoU equation [41]. ........................................................................ 47

20. The higher the IoU, the better the performance [24]. ....................................................... 48

21. (a) Object detection algorithm after 10 epochs (b) Object detection algorithm for an

image without any link (c) Improved Object detection algorithm after 200 epochs. ....... 49

22. Accuracy between two Optical Character Recognition techniques and our model on ten

im-ages using regular font style. ....................................................................................... 50

23. Regular Expression used to recognize a URL in a string of text. ..................................... 50

24. Our model experiencing overfitting. ................................................................................. 51

25. OCR image conversion to text versus our object detection model. .................................. 52

26. Example italicized text. ..................................................................................................... 53

27. Detection Accuracy between Optical Character Recognition techniques and our model on

ten images using italicized font style. ............................................................................... 53

28. Detection on different input image sizes. ......................................................................... 54

29. Our models performance on slightly varying font size. .................................................... 58

 8

Chapter I: Introduction

Overview

In this chapter, we introduce the background and focus area of our thesis. We highlighted

our motivations, scope, and thesis organization. We will explore YOLO’s potential in recogniz-

ing objects and creating bounding box around them.

Motivation

Image recognition/image processing is in the forefront of Artificial Intelligence today. It

is however far from perfection. Seemingly simple scenarios, such as object detection, face recog-

nition, removing motion blur, etc. and more complex scenarios such as compression artefacts,

scratch detection, sensor noise, and spilling detection are applications of image recognition/im-

age processing.

Digitized images are often represented as a two-dimensional (2D) array of pixels values.

Each pixel value which makes up the color scheme of the image is often influenced by an array

of factors such as light intensity. Visual scene is projected unto a surface, where receptors (natu-

ral or artificial) produce values that depend on the intensity of incident light.

These exciting concepts are however hard to implement. Forming an image leads to loss

of details of information while collapsing a three-dimensional (3D) image into a two-dimen-

sional image. Many other factors are responsible for why image recognition/ image processing is

hard. Some of such factors are noise in the image (pixels values that are off from its surrounding

pixels), mapping from scene to image etc.

In recent years, during the ImageNet Large Scale Visual Recognition Competition

(ILSVRC, 2015), computers were going better than humans in the image classification task [35].
 9

In 2016, a faster object detector, YOLO, was proposed to implement object detection in real-time

situation. Our motivation is to apply YOLO to object detection task of URL links within an im-

age scene. We will also be comparing the speed and accuracy of this with an OCR software.

Research aim

The research aim of this thesis is design and implement an AI algorithm to detect URL

links within image scenes using YOLO. Subsequently, we will use an Optical Character Recog-

nition (OCR) software to convert these images to text and use regular expression to search for

links within the text. Both approaches will be compared for accuracy. We also proposed a few

improvements and future work later.

Problem description and Research Questions

The first research question of this study is to verify whether YOLO is a good model for

general object detection or not. Furthermore, given the nature of URLs and the fact that this class

of classification does not yet exist, whether YOLO is a potential candidate for URL detection

will be considered.

The second research question is to evaluate the accuracy of our YOLO detection model

and the deep learning optical character recognition model, pytesseract.

What is an image in Computer Science?

An image (from Latin: imago), is an artefact that depicts visual perception, such as a

photograph or other two-dimensional picture, that resembles a subject - usually a physical object

- and thus provides a depiction of it [1]. An image is a visual representation of something [2]. Ac-

cording to The Merriam-Webster Dictionary, an image is.

 10

“a: a visual representation of something: such as

(1): a likeness of an object produced on a photographic material

(2): a picture produced on an electronic display (such as a television or computer screen)

b: the optical counterpart of an object produced by an optical device (such as a lens or mirror)

or an electronic device” [3]

While the orthodox idea of an image is that it usually only two dimensional and thus must

appear on a screen, the technical visage of the word suggests that an image can also be three-di-

mensional. Thus, images may be two-dimensional, such as a photograph or screen display, or

three-dimensional, such as statues or holograms. They may be captured by optical devices – such

as cameras, mirrors, lenses, telescopes, microscopes, etc. and natural objects and phenomena,

such as the human eye or water.

More interestingly perhaps, is that the term image lends itself to more than physical ap-

plications. The term also lends itself to mental and metaphysical uses. As such, an image can be

interpreted to mean, a mental image - an impression created in the imagination of a person upon

the suggestion of a particular concept, and a divine image- an image believed to be transmitted

into the human psyche by a supernatural entity. An image may also be classified as being either

fixed or volatile. A volatile image is one that exists only for a short period of time. This may re-

flect an object by a mirror, a projection of a camera obscura, or a scene displayed on a cathode

ray tube. A fixed image, also called a hard copy, is one that has been recorded on a material ob-

ject, such as paper or textile by photography or any other digital process [4].
 11

The term ‘image’ in Computer Science is employed in a more technical scope. It is both

restricted as well as specific in its application. Image processing for instance is a set of computa-

tional techniques for analyzing, enhancing, compressing, and reconstructing images. Its main

components are importing, in which an image is captured through scanning or digital photog-

raphy; analysis and manipulation of the image [5], In information technology, the term has sev-

eral usages.

Firstly, it may be employed in its base form, in the way, an image is a picture that has

been made or copied and stored in electronic form. An image can also be described in terms of

vector graphics or raster graphics. An image stored in raster form is sometimes called a bitmap.

Bitmap images are made from small parts called pixels. Vector images are made using coordi-

nates and geometry. Images can be compressed to reduce file size [6].

Common image file formats online include:

1. JPEG (pronounced JAY-peg) is a graphic image file produced in compliance to the

standard set by the Joint Photographic Experts Group, an ISO/IEC group of experts that

develops and maintains standards for a suite of compression algorithms for computer im-

age files. JPEGs usually have a .jpg file extension.

2. GIF (pronounced JIF by many, including its designer; pronounced GIF with a hard G

by many others) stands for Graphics Interchange Format. The GIF employs the use of 2D

raster data type and is encoded in binary. GIF files ordinarily have the .gif exten-

sion.GIF89a is an animated GIF image, formatted according to GIF Version 89a. One of

the merits of this format is that it possesses the ability to create an animated image that

can be played after transmitting to a viewer page that moves - for example, a twirling
 12

icon or a banner with a hand that waves or letters that magically get larger. A GIF89a can

also be specified for interlaced GIF presentation.

3. PNG (pronounced ping) is an acronym for Portable Network Graphics. It is a file for-

mat used for image compression and was designed to provide a number of enhancements

over the GIF format. Like a GIF, a PNG file is compressed in lossless fashion, this means

that all image information is restored when the file is decompressed during viewing. Files

typically have a .png extension.

4. SVG is Scalable Vector Graphics, the description of an image as an application of

XML. Any program such as a browser that recognizes XML can display the image using

the information provided in the SVG format. Scalability means that the file can be

viewed on a computer display of any size and resolution, whether the small screen of a

smartphone or a large widescreen display in a PC. Files usually have .svg extension.

5. TIFF (Tag Image File Format) is a common format for exchanging raster graphics

(bitmap) images between application programs, including those used for scanner images.

A TIFF file can be identified as a file with a .tiff or ".tiff" file name suffix.

Furthermore, the term is also often used to mean a disk image, a disk image is a copy of

the entire contents of a storage device, such as a hard drive or DVD. The disk image represents

the content exactly as it is on the original storage device, including both data and structure infor-

mation.

Lastly, another use of the term image is for a section of random-access memory (RAM)

that has been copied to another memory or storage location [2].

 13

What is Object Detection?

Object detection is a computer vision implementation that makes a system (an algorithm)

about to estimate the location of objects in a digitized scene such as an image or video. Usually, a

bounded box is wrapped around the detected object which helps humans locate the object

quicker than unprocessed images. For this discourse, an object is the representation of a physical

object (URL) in an image. In image processing, it is an identifiable portion of an image that can

be interpreted as a single unit [7]. This creates a sharp contrast to the layman’s idea that an image

or an object are interchangeable.

Usually, an image may contain one or more objects, the discernibility of which is of up-

most importance. For instance, in a single image the objects contained can range from a single

unit to as many objects of as numbers, bordering on infinity.

Although “detection” could mean locating a hidden concealed object, detection may also

mean the ability of an intelligence to signify the existence and identification of an object. The ob-

ject in question does not have to be hidden. This later form is the form in which we based this

thesis.

Interpreting the object localization for object detection can be done in various ways, in-

cluding creating a bounding box around the object or marking every pixel in the image which

contains the object (called segmentation) [8]. Thus, given an image or video stream, an object

detection model should be able to identify which of a known set of objects might be present and

provide information about their positions within the image [9].

Object Detection is widely employed in various computer vision tasks such as image an-

notation, activity recognition, face detection, face recognition, video object co-segmentation. It is
 14

also used in tracking objects. For example, tracking a ball during a football match, tracking

movement of a cricket bat, or tracking a person in a video are some of its various uses [10].

Fundamentally, two approaches to image detection exist, they are machine learning-based

approaches and deep learning-based approaches [11]. In more traditional ML-based approaches,

computer vision techniques are employed to analyze various features of an image, such as the

color histogram or edges, to identify groups of pixels that may belong to an object. These fea-

tures are then inputted into a regression model that predicts the location of the object along with

its label [11].

Some Machine learning approaches are Viola–Jones object detection framework based on

Haar features, Scale-invariant feature transform (SIFT) and Histogram of oriented gradients

(HOG) features [10]. On the other hand, deep learning-based approaches employ convolutional

neural networks (CNNs) to perform end-to-end, unsupervised object detection, in which features

do not need to be defined and extracted separately [11]. Some Deep learning approaches are: Re-

gion Proposals (R-CNN,Fast R-CNN,Faster R-CNN, cascade R-CNN.), Single Shot MultiBox

Detector (SSD), You Only Look Once (YOLO), Single-Shot Refinement Neural Network for

Object Detection (RefineDet),Retina-Net and Deformable convolutional networks.

What is Image Recognition?

Object detection is often confused with image recognition (Figure 1). A picture of a dog

receives the label “dog”. A picture of two dogs, still receives the label “dog”. Object detection,

on the other hand, draws a box around each dog and labels the box “dog”. The model predicts

where each object is and what label should be applied. In that way, object detection provides

more information about an image than recognition [11].

 15

Recognition in this context is the ability of an intelligent system to identify an object

based on certain similarities that it shares with another object that the intelligence has previously

encountered. Recognition may be based on inference or relation, that is, a situation whereby an

intelligence is able to recognize an object because it recognizes similarities in form and proper-

ties. Recognition may also occur because the Artificial intelligence has encountered the exact

specimen at a previous instance.

In human beings’ recognition is a cognitive process that happens seamlessly and almost

instantly without any hitch. The human brain is capable of learning and adapting information

with minimal effort such that even humans that are still in the developmental stages of their ex-

istence can recognize objects and patterns easily. Humans can recognize a multitude of objects in

images with little effort, even though the image of the objects may vary somewhat in different

viewpoints, in many different sizes and scales or even when they are translated or rotated. Ob-

jects can even be recognized when they are partially obstructed from view [12]. Artificial intelli-

gence, however, does not innately possess this cognitive ability. For Artificial Intelligence to ac-

quire this level of skill they must acquire training. This training is usually acquired by ‘teaching’

the A.I. using coding, datasets, and databases. This task is still a challenge for computer vision

systems given these A.I. systems need to be trained for each class of object it is meant to recog-

nize.

The object recognition problem can be defined as a labelling problem based on models of

known objects. Formally, given an image containing one or more objects of interest (and back-

ground) and a set of labels corresponding to a set of models known to the system, the system

should be able to accurately assign correct labels to regions, or a set of regions, in the image. The
 16

object recognition problem is closely tied to the segmentation problem: without at least a partial

recognition of objects, segmentation cannot be done, and without segmentation, object recogni-

tion is not possible [13].

Object recognition is an extremely difficult computational problem. The core problem is

that each object in the world can cast an infinite number of different 2-D images onto the retina

as the object's position, pose, lighting, and background vary relative to the viewer. Yet the brain

solves this problem effortlessly. Progress in understanding the brain's solution to object recogni-

tion requires the construction of artificial recognition systems that ultimately aim to emulate our

own visual abilities, often with biological inspiration [14].

Figure 1. Image Recognition and Object Detection difference [36].

 17

Chapter II: Background

Overview

In this Chapter, we will review the basic concepts of general neural networks, explain a

subclass of neural network, Convolutional Neural Network (CNN), Batch Normalization, Max

Pooling, Activation Function, and Rectified Linear Unit (RELU). We will also describe a specific

CNN based object detection model, YOLO, which is the most accurate real-time object detector

[35].

The neural networks described in this thesis are the feed-forward neural networks, in

which information is transmitted in a feedforward manner.

Image Processing

Image processing is a technique that is used to detect objects/patters and perform some

operations on an image usually with the aim to retrieve value information from the image or en-

hance it. The processing of digital images can be divided into several classes: image enhance-

ment, image restoration, image analysis, and image compression [36].

Image Enhancement: This is a technique used to adjust digital images for better suitable

displays or analysis. An image can be enhanced using median filtering, linear contrast adjust-

ment, histogram equalization, etc.

Figure 2. Enhancing grayscale images with histogram equalization [37].

 18

Image Restoration: Restoration is usually performed on burry or noisy images. Images

can be blurred due to many factors like relative motion the camera and the object. Noises are of-

ten caused by environmental conditions such as rain, snow, etc. or even by thermal signals [36].

Image Analysis: Images that contain useful information can be used to outline objects

and describe them. Some examples of image analysis are edge extraction, image segmentation,

and texture and motion analysis [36].

Image Compression: When an image is compressed, the image byte is minimized with-

out degrading the quality of the image. The advantage of compression is more images can be

stored on a disk or memory after compression [36].

Artificial intelligence and learning

The term artificial intelligence was first used in 1956, at a computer science conference

in Dartmouth. It described an attempt to digitally model how the human mind worked and based

on this knowledge create more advanced computers. Even though some truly groundbreaking

progress has been made, today, artificial intelligence is ubiquitous, but computers are still far

from modelling human intelligence to perfection. The goal of the “Strong” AI is to become artifi-

cial persons: machines that have all the mental powers we have, including phenomenal con-

sciousness. “Weak” AI, on the other hand, seeks to build information-processing machines that

appear to have the full mental repertoire of human persons (Searle 1997). Weak, or narrow AI on

the other hand is created to be quite adept at carrying out the purpose that it was defined for, but

it will not pass for human in any other field outside of its defined capacities. An example of this

is Deep Blue, the first computer to defeat a human in chess. It was able to defeat the famous

chess master Garry Kasparov in 1997.

 19

To facilitate this growth and development in the field, machines must learn how to be-

have and this process if facilitated through a system known as machine learning.

Machine Learning

Machine learning is an area within Artificial Intelligence (AI), that focuses on decision

making and predictions. The primary aim is to allow the computer to be further developed with-

out any human intervention, it will be trained based on observations and data [36]. Machine

learning algorithms are often categorized as supervised or unsupervised, where supervised algo-

rithms apply what has been learned in the past to new data using labeled examples. This type of

problem can be either predicting a continuous quantity (Regression) or discrete class labels

(Classification). Unsupervised learning includes grouping a set of uncategorized data by finding

structures or patterns (Clustering) or finding relationships between variables in big data (Associ-

ation).

Machine learning is a subset of the larger field of Artificial Intelligence that “focuses on

teaching computers how to learn without the need to be programmed for specific tasks [15],” “In

fact, the key idea behind ML (sic) is that it is possible to create algorithms that learn from and

make predictions on data [15].” To educate a machine. Three important components must be pre-

sent, namely, a dataset, features, and an algorithm. These components will be discussed serially

in subsequent paragraphs.

A dataset is a pool of samples compiled with the intention of educating a machine on how

to go about a task. These pools can sometimes be numbers, images, texts, or any other kind of

data. It can sometimes be quite strenuous and expensive to compile a data set depending on how
 20

specialized the task is expected to be. For instance, a data set of exclusive Rolex watches would

be easier to compile than a data set of all types of cars in existence.

Features are important pieces of data that act as the key to the solution of the task. They

show the machine what to look for in an image. The way features are defined vary based on what

the machine is expected to look out for. For instance, to predict the price of an apartment, it

would prove to be too difficult to prove by way of linear regression the cost of the property based

on the combination of its length and width However, it may be easier to predict a price based on

the correlation between the price and location of a building. Therefore, there is no gainsaying the

fact that the accuracy of the features provided affects the eventual performance of the machine.

Which is what occurs when the machine is trained with labeled data which contain the “right so-

lutions”, and a validation set. During the learning process, the program is expected to get the

“right” solution. And then, the validation set is used to tune hyperparameters to avoid overfitting.

However, in unsupervised learning, features are learned with unlabeled input data. In his case the

machine is not given any features to familiarize itself with, it learns to notice the patterns by it-

self.

An algorithm is a finite sequence of well-defined, computer-implementable instructions,

typically to solve a class of problems or to perform a computation [16]. Algorithms are always

unambiguous and are used as specifications for performing calculations, data processing, auto-

mated reasoning, and other tasks [17]. The instructions describe a computation that, when exe-

cuted, proceeds through a finite number of well-defined successive states, eventually producing

“output”.
 21

Machine Learning Evaluation

Evaluating a machine learning model requires that our dataset be split appropriately. It is

typically to have 70% training dataset and 30% for testing. However, it is also advisable to have

about 20% validation dataset to make it possible to evaluate the model while still building and

tuning the model. This helps to avoid overfitting and discourages using the training dataset for

testing. After shuffling the data, a split of 60% training dataset, 20% validation dataset, and 20%

testing dataset is advisable.

Speed, precision, recall, and accuracy are important metrics of a machine learning algo-

rithm to be evaluated. The accuracy of an object detection model is calculated using the Mean

Average Precision (mAP) [36].

Accuracy is the ratio of correct predictions to the total number of input samples.

correct predictions
𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
all predictions

Precision is the ratio of relevant examples (true positives) among all the examples which

are predicted to belong in a certain class [38].

true positives
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
true positives + false positives

Recall is the ratio of examples which were predicted to belong to a class with respect to

all the examples that truly belong in that class [38].

true positives
𝑟𝑒𝑐𝑎𝑙𝑙 =
true positives + false negatives

True Positive (TP) [36]: The prediction that was made is true, and the actual output is

positive.
 22

False Positive (FP) [36]: The prediction that was made is false, and the actual output is

positive.

False Negatives (FN) [36]: The prediction that was made is false, and the actual output is

negative.

Deep Learning

Apart from the machine learning earlier discussed there are new evolved specialized

forms of learning which have grown apart and form a subset of machine learning known as Deep

Learning. Deep learning is a more effective branch of machine learning based off the perceived

structure of the human brain, Deep learning algorithms use complex multi-layered neural net-

works, where the level of abstraction increases gradually by non-linear transformations of input

data.

In a neural network, information moves from one layer to another over a set of over-con-

necting channels. These channels are called weighted channels due to the amount of value at-

tached to them.

All neurons possess a unique number known as bias. This bias combined with the

weighted sum of inputs reaching the neuron is then applied to the activation function. The result

of the function determines if the neuron gets activated. Every activated neuron passes on infor-

mation to the following layers. This continues up to the second last layer. The output layer in an

artificial neural network is the last layer that produces outputs for the program [18]. The field of

deep learning is very useful in fields such as object recognition as well as voice recognition.

In 2013, R-CNN was the first algorithm to apply deep learning to Object detection. It out-

performed the previous algorithms by up to 30% on the Visual Object Classes Challenge
 23

(VOC2012). This affirmed its place as the frontrunner of the object detection industry. It beats

the previous ones by more than 30% on the VOC2012 and was therefore a huge improvement in

the fields of Object detection.

Object Detection involves two difficulties: locating objects and suppling the appropriate

classifications.

Neural Network

Neural network is a computer system modeled on the human brain and the nervous sys-

tem. Wikipedia defines it as a network of circuit of neurons, or in a modern sense, an artificial

neural network, composed of artificial neurons or nodes. Thus, a neural network is either a bio-

logical neural network, made up of real biological neurons, or an artificial neural network, for

solving artificial intelligence (AI) problems.

Neuron

Neuron, in biology, is the basic working unit of the brain, an electrically excitable cell

that communicates with other cells via specialized connections called synapses. A neuron is the

atomic element of a neural network. A perceptron is the mathematical model of a biological neu-

ron where electric signals are represented as numerical values.

Figure 3. Structure of Perceptron [35].

 24

Figure 4. Output of a Perceptron [35].

A perceptron has many inputs and one output. It calculates the weighted sum of the input

values to represent the total strength of the input signals and applying a step function as the acti-

vation function on the sum to output the result. The output generated is fed into other perceptron

and the final output is bounded between 0 and 1. A multi-layered perceptron is called neural net-

works.

Figure 5. Activation functions. (1) the left curve is a sigmoid function curve. (2) the right curve
is a tanh function curve [35].
Structure of Neural Networks

Neural networks typically consist of multiple layers and each layer has many neurons as

shown in Figure 6. The first layer is known as the input layer where each circle represents an in-

put neuron. The rightmost layer is the output layer and contains output neuron(s), and all the
 25

middle layers are hidden layers. General neural networks have fully connected neurons and neu-

rons between adjacent layers are connected for passing information. For example, if two adjacent

layers have 𝑚 and 𝑛 neurons respectively, the total number of connections will be 𝑚 × 𝑛 [35].

Figure 6. A sample fully connected neural network with only one hidden layer [35].
Cost Function and Training

Training involves fine-tuning a bad performing neural network (due to randomly initial-

ized weights) into a network with high accuracy. Training is equivalent to minimizing loss func-

tion. Neural networks learn from the errors between the predictions and ground truths by updat-

ing its parameters (weights) [35]. A cost function is a measure of error between the predicted

value of a model and what the actual value is. The error range continues to shrink during the

training process until the value stays constant which indicates the training process is complete.

Figure 7 shows a visual representation of a training process. The bowl-shaped error surface indi-

cates the error sliding from the edge to the bottom (where the minimum error is) [35].
 26

Figure 7. Visualization of a Training Process [35].

Convolutional Neural Network (CNN)

Convolutional Neural Network (CNN) first surfaced around 1998 [35] but was proved to

be an efficient tool for image classification in 2012 at the Large Scale Visual Recognition Chal-

lenge. CNN is a deep learning algorithm which can take an input image, assign importance

(learnable weights and biases) to various aspects/objects in the image and be able to differentiate

one from the other. The pre-processing required in a CNN is much lower as compared to other

classification algorithms. While in primitive methods, filters are hand-engineered, with enough

training, CNN can learn these filters/characteristics [39].

 27

Pooling

After each convoluted operation, pooling operation follows to further simplify the infor-

mation. The pooling process simply compresses previous feature maps into condensed feature

maps. The two most common pooling methods are average-pooling and max-pooling. Average-

pooling means to output the average value among the pooling filed, and max-pooling is to select

the maximum value. Figure 8 shows a max pooling operation with a 2 × 2 pooling filter. The

size of the feature map is reduced from 24 × 24 to 12 × 12. We will be using max-pooling for

our experiment [35].

Figure 8. Max-pooling. Pooling from 24 × 24 to 12 × 12 [35].

General Object Detection Model

Figure 9 is an object detection structure which has a region proposal component followed

by a CNN classifier. Researchers use region proposal methods to produce a bunch of candidate

regions, each of which may contain one kind of object. Each region is then passed through the
 28

CNN classification algorithm. The model is to convert a multiple object detection problem into a

single object classification problem. However, these region proposal methods are much slower

than classification part, which becomes the bottleneck for the whole system. The drawback of

this structure is that we cannot tradeoff accuracy for detection speed in time critical applications

[35].

Figure 9. Structure of a CNN Based Object Detection Model [35].

Unified Detection Model – YOLO

The YOLO model was first brought into existence by Joseph Redmon in his paper “You

only look once, Unified, Real-time object detection”. The mechanism for the algorithm employs

the use of a single neural network that takes a photograph as an Input and attempts to predict

bounding boxes and class labels for each bounding box directly. Although this offered less pre-

dictive accuracy, which was mostly due to more localization errors, it boasted speeds of up to 45

frames per second and up to 155 frames person on speed optimized versions of the model [19].

“Our unified architecture is extremely fast. Our base YOLO model processes images in

real-time at 45 frames per second. A smaller version of the network, Fast YOLO, processes an

astounding 155 frames per second …” [21]

 29

To begin with the model operates by splitting the inputted image into a grid of cells,

where each cell is responsible for predicting a bounding box if the center of a bounding box falls

within it. Each grid cell predicts a bounding box involving the x, y coordinate and the width and

height and a metric of valuation of quality known as a confidence score. A class prediction is also

based on each cell.

To supply more emphasis an instance will be provided. For example, an image may be

divided into a 7 × 7 grid and each cell in the grid may predict 2 bounding boxes, resulting in 94

proposed bounding box predictions. The class probabilities map and the bounding boxes with

confidences are then combined into a final set of bounding boxes and class labels.

The YOLO was not without shortcomings, the algorithm had a number of limitations be-

cause of the number of grids that it could run on as well as some other issues which will be ad-

dressed subsequently. Firstly, the model uses a 7 × 7 grid and since each grid can only identify

an object, the model restricts the maximum number of objects detectable to 49. Secondly, the

model suffers from what is known as a close detection model, since each grid is only capable of

detecting one object, if a grid cell contains more than one object it will be unable to detect it.

Thirdly, a problem might arise because the location of an object might be more than a grid, thus,

there exists a possibility that the model might detect the object more than once [20]. Due to the

aforementioned problems encountered when running YOLO, it was fairly obvious that localiza-

tion error and other problems of the system needed to be addressed. As a result of that, YOLOv2

was created as an improvement to deal with the issues and questions posed by its predecessor.

Therefore, localization errors as well as errors of real were significantly addressed in the new
 30

version. The model was updated by Joseph Redmon and Ali Farhadi to further revamp model

performance in their 2016 paper named “YOLO9000: Better, Faster, Stronger [19]”.

Structure of YOLO

YOLO is implemented as a convolution neural network and has been evaluated on the

PASCAL VOC detection dataset. It consists of a total of 24 convolutional layers followed by 2

fully connected layers. The layers are separated by their functionality in the following manner:

1. First 20 convolutional layers followed by an average pooling layer and a fully

connected layer is pre-trained on the ImageNet 1000-class classification dataset.

2. The pretraining for classification is performed on dataset with resolution

224 × 224.

3. The layers comprise of 1 × 1 reduction layers and 3 × 3 convolutional layers.

4. Last 4 convolutional layers followed by 2 fully connected layers are added to train

the network for object detection.

5. Object detection requires more granular detail hence the resolution of the dataset

is bumped to 448 × 448.

6. The final layer predicts the class probabilities and bounding boxes.
 31

Figure 10. YOLO Structure [35].

The final layer uses a linear activation whereas the other convolutional layers use leaky

ReLU activation.

The input is 448 × 448 image, and the output is the class prediction of the object en-

closed in the bounding box.

Network Structure

Figure 11. YOLO Network Architecture [42].

The whole system can be divided into two major components: Feature Extractor and De-

tector; both are multi-scale. When a new image comes in, it goes through the feature extractor
 32

first so that we can obtain feature embeddings at three (or more) different scales. Then, these fea-

tures are feed into three (or more) branches of the detector to get bounding boxes and class infor-

mation [42].

Intersection over Union (IoU)

YOLOs default metric for measuring overlap between two bounding boxes or masks is

Intersection over union (𝐼𝑜𝑈). Any algorithm that provides predicted bounding boxes as output

can be evaluated using 𝐼𝑜𝑈 [20]. If the prediction is completely correct, 𝐼𝑜𝑈 = 1. The lower the

𝐼𝑜𝑈, the worse the prediction result.

To apply Intersection over union, certain parameters must be met to evaluate an (arbi-

trary) object detector, they are: The ground-truth bounding boxes these are the hand labeled

bounding boxes from the testing set that specify where in the image our object is and the pre-

dicted bounding boxes from our model. If these two sets of bounding boxes are present it is pos-

sible to apply Intersection over union. Thus, computing Intersection over Union can therefore be

calculated as; The area of overlap divided by the area of union [22].

Plainly put, the intersection over union is a ratio of the similarity of the ground truth

bounding box and the predicted bounding box, thus predicting a rough estimate on how much an

Artificial intelligence can rely on the predictions made by the algorithm. This is an improvement

over binary models which label predictions as either correct or incorrect. Also due to the variance

in parameters between the model and the object it is quite unrealistic to have a 100% match be-

tween the (𝑥, 𝑦) coordinates a predicted bounding box and the (𝑥, 𝑦) coordinates of the ground

truth box [22]. Thus, this equation ensures that boxes with a larger area of overlap get higher

scores than those with lesser areas thus cementing Intersection over union as excellent metric for
 33

evaluating custom object detectors. Generally, any score greater than 0.8 is a good score [23].

This application of intersection over union is used during the testing phase at the completion of

the training [24].

Another interesting application of this model occurs where there is more than one bound-

ing box for the same object. the metric helps to eliminate bounding boxes with lesser scores.

How this is done is that if there are two bounding boxes with very high confidence scores, then

the chances are that the boxes are detecting the same object therefore the box with the higher

confidence rating is selected. However, where confidence rating of the two boxes is low, it is

likely that they are predicting separate objects of the same class, such as two different cars in the

same picture. This application is used after the model has completed training and is being de-

ployed [24].

As earlier mentioned, Intersection over union is a good metric for evaluating the quality

of a task. Before further elucidation, it is pertinent to define some terminologies first, these ter-

minologies are true positive, true negatives, false negatives, and false positives. Simply put, a

true positive is any correctly drawn annotation with a value greater than 0.5, a true negative is

when an F1 refuses to draw any annotation because there simply is not one to be drawn. There is

no value here because no annotation is drawn, thus there is no way to calculate true negatives. A

false negative occurs where there are missing annotations while a false positive occurs where

there are incorrectly drawn annotations that have an 𝐼𝑜𝑈 score of less than 0.5. An equation

known as accuracy is usually used to measure the performance of a task because it is an incredi-

bly straight forward measurement as well as for its simplicity. it is simply a ratio of correctly

drawn annotations to the total expected annotations (ground truth). Thus, it is calculated as being
 34

equal to the sum of True positive and True negative divided by the sum of True positive, False

positive, True negative and False negative.

Direct Location Prediction

In the early versions of the YOLO model, there were no constraints on the location pre-

diction. The predicted bounding box was not tethered therefore it could occur far from the origi-

nal grid location. This anomaly resulted in a very unstable model [25]. The bulk of the instability

resulted from predicting the (𝑥, 𝑦) locations for the box. In region proposal networks the network

predicts values 𝑡𝑥 and 𝑡𝑦 and the (𝑥, 𝑦) center coordinates are calculated as:

𝑥 = (𝑡𝑥 ∗ 𝑤𝑎 ) − 𝑥𝑎

𝑦 = (𝑡𝑦 ∗ ℎ𝑎 ) − 𝑦𝑎

For example, a prediction of tx = 1 would shift the box to the right by the width of the an-

chor box, a prediction of 𝑡𝑥 = −1 would shift it to the left by the same amount [25]. This for-

mulation did not exist within any boundaries and as such, any anchor box could end up at any

point in the image regardless of what location predicted the box. with random initialization it

took the model an obscene amount of time to stabilize to predict sensible offsets [25]. The subse-

quent versions of YOLO diverged from this approach and devised a means to properly tackle the

situation. YOLOv2 bounds the location using logistic activation.

Sigma (𝜎), which ensures that the value remains between 1 and 0 [26]. Given the anchor

box of size (𝑝𝑤 , 𝑝ℎ ) at the grid cell with its top left corner at (𝑐𝑥 , 𝑐𝑦 ), the model predicts the off-

set and the scale, (𝑡𝑥 , 𝑡𝑦 , 𝑡𝑤 , 𝑡ℎ ) and the corresponding predicted bounding box has center

(𝑏𝑥 , 𝑏𝑦 ) and size (𝑏𝑤 , 𝑏ℎ ). The confidence score is the sigmoid (𝜎) of another output to. Since
 35

the location prediction is constrained, the parameterization is easier to learn, making the network

more stable. The employment of dimension clusters along with directly predicting the bounding

box’s center location improves YOLO by almost 5% over the version with anchor boxes n, m.

Bounding Box Prediction

A bounding box is a relatively closed space within which points, objects, or a group of

objects may be contained. For YOLO, the cell in which the center of an object resides, is the cell

responsible for detecting that object. Each cell will predict B bounding boxes and a confidence

score for each box. The confidence score will be from ‘0.0’ to ‘1.0’, with ‘0.0’ being the lowest

confidence level and ‘1.0’ being the highest; if no object exists in that cell, the confidence scores

should be ‘0.0’, and if the model is completely certain of its prediction, the score should be ‘1.0’.

These confidence levels capture the model’s certainty that there exists an object in that cell and

that the bounding box is accurate. Each of these bounding boxes is made up of 5 numbers: the x

position, the y position, the width, the height, and the confidence. The coordinates ‘(x, y)’ repre-

sent the location of the center of the predicted bounding box, and the width and height are frac-

tions relative to the entire image size. The confidence represents the Intersection over Union

(IoU) between the predicted bounding box and the actual bounding box, referred to as the ground

truth box [27]. In general, there are five types of bounding boxes, i.e., a surrounding sphere (SS),

an axis-aligned bounding box (AABB), an oriented bounding box (OBB), a fixed-direction hull

(FDH), and a convex hull (CH) [29]. The AABB refers to a box whose axis is parallel to the co-

ordinate axis. It is the rectangle formed by selecting the maximum and minimum horizontal and
 36

vertical coordinates in each vertex of the two-dimensional shape and is one of the most com-

monly used bounding box types [29]. Bounding box functions to mark the object area that has

been detected [28].

In digital image processing, the bounding box is the coordinates of a rectangle wishing

which an object may be contained when it is placed over a page, a canvas, a screen, or any other

similar bi-dimensional background [30]. In the field of object detection, a bounding box is usu-

ally used to describe an object location. The bounding box is a rectangular box that can be deter-

mined by the 𝑥 and 𝑦 axis coordinates in the upper-left corner and the 𝑥 and 𝑦 axis coordinates in

the lower-right corner of the rectangle [31]. The first version of YOLO directly predicted all four

values which describes a bounding box. the x and y coordinates of each bounding box are de-

fined relative to the top left corner of each grid cell and normalized by the cell dimensions such

that the coordinate values are bounded between 0 and 1. However in YOLOv2 there was a shift

in paradigm and the algorithm employed dimensional clusters in place of anchor boxes, 4 coordi-

nates are predicted for each bounding box, 𝑡𝑥 , 𝑡𝑦 , 𝑡𝑤 , 𝑡ℎ , If the cell is offset from the top left cor-

ner of the image by (𝑐𝑥 , 𝑐𝑦 ) and the bounding box prior has width and height 𝑝𝑤 , 𝑝ℎ , then the

predictions correspond to [32]:

Figure 12. Bounding Box prediction formula [32].

 37

During training, the sum of squared error loss is used Assuming that the ground truth for

some coordinate prediction is 𝑡̂∗ our gradient is the ground truth value (computed from the

ground truth box) minus our prediction: 𝑡̂∗ − 𝑡∗ . This ground truth value can be easily computed

by inverting the equations above. YOLOv3 predicts an objectness score for each bounding box

using logistic regression. This should be 1 if the bounding box prior overlaps a ground truth ob-

ject by more than any other bounding box prior is not the best but does overlap a ground truth

object by more than some threshold the prediction is ignored. We use the threshold of .5. usu-

ally, a system only assigns one bounding box prior for each ground truth object. If a bounding

box prior is not assigned to a ground truth object it incurs no loss for coordinate or class predic-

tions, only objectness [33].

The concept of a bounding box prior was introduced in YOLOv2. Previously, the model

was expected to provide unique bounding box descriptors for each new image, a collection of

bounding boxes is defined with varying aspect ratios which embed some prior information about

the shape of objects we are expecting to detect. Redmon offers an approach towards discovering

the best aspect ratios by doing k-means clustering (with a custom distance metric) on all of the

bounding boxes in the training dataset [33].

Thus, instead of predicting the bounding box dimension directly, the task is reformulated

to simply predict the offset from the bounding box prior in order to fine-tune the predicted

bounding box dimensions. The result of which is that it makes the prediction task easier to learn.

Objectness (and assigning labeled objects to a bounding box)

“objectness” score 𝑝𝑜𝑏𝑗 is trained to approximate the intersection over union between the

predicted box and the ground truth label. When the loss during training is calculated, Objects are
 38

matched to whichever bounding box prediction on the same grid cell produced the highest 𝐼𝑜𝑈

score. For unmatched boxes, the only descriptor which will be included in the function is 𝑝𝑜𝑏𝑗 .

Upon the introduction of additional bounding box priors in YOLOv2, it was possible to

assign objects to whichever anchor box on the same grid cell has the highest IoU score with the

labeled object.

YOLO (version 3) redefined the "objectness" target score 𝑝𝑜𝑏𝑗 to be 1 for the bounding

boxes with highest 𝐼𝑜𝑈 score for each given target, and the score 0 for all remaining boxes.

However bounding boxes which have a high 𝐼𝑜𝑈 score above a defined threshold but not the

highest score when calculating the loss will not be included. This simply means that it does not

produce the most appropriate prediction because it is not the best possible prediction [33].

Class Labels

YOLO (version 3) uses sigmoid activations for multi-label classification, noting that

SoftMax (from the previous versions) is not necessary for good performance. This choice will

depend on your dataset and whether or not your labels overlap (e.g., "golden retriever" and

"dog").

Output Layer

YOLO (version 3) has 3 output layers. These output layers predict box coordinates at 3

different scales. The output prediction is of the form 𝑤𝑖𝑑𝑡ℎ × ℎ𝑒𝑖𝑔ℎ𝑡 × 𝑓𝑖𝑙𝑡𝑒𝑟𝑠.

YOLO (version 3) replaced the skip connection splitting for a more standard feature pyr-

amid network output structure. With this method, there is an alternate between outputting a pre-

diction and up sampling the feature maps (with skip connections). This allows for predictions
 39

that can take advantage of finer-grained information from earlier in the network, which helps for

detecting small objects in the image [33].

Pytesseract

Tesseract is an open-source text recognition (OCR) Engine, available under the Apache

2.0 license. It can be used directly, or (for programmers) using an API to extract printed text from

images. It supports a wide variety of languages. Tesseract does not have a built-in graphics user

interface (GUI) [43]. Pytesseract is the tesseract framework for python language.

Legacy Tesseract 3.x was dependent on the multi-stage process where we can differenti-

ate steps into word finding, line finding, and character classification. Word finding was done by

organizing text lines into blobs, and the lines and regions are analyzed for fixed pitch or propor-

tional text. Text lines are broken into words differently according to the kind of character spac-

ing. Recognition then proceeds as a two-pass process. In the first pass, an attempt is made to rec-

ognize each word in turn. Each word that is satisfactory is passed to an adaptive classifier as

training data. The adaptive classifier then gets a chance to, more accurately, recognize text lower

down the page [43].

Modernization of the Tesseract tool was an effort on code cleaning and adding a new

Long Short-Term Memory (LSTM) model. The input image is processed in boxes (rectangle)

line by line feeding into the LSTM model and giving output.

After adding a new training tool and training the model with a lot of data and fonts, Tes-

seract achieves better performance. Still, not good enough to work on handwritten text and weird

fonts. It is possible to fine-tune or retrain top layers for experimentation [43].

 40

Preprocessing for Tesseract

Tesseract does various image processing operations internally (using the Leptonica li-

brary) before doing the actual OCR. It generally does a very good job of this, but there will inev-

itably be cases where it is not good enough, which can result in a significant reduction in accu-

racy. Pre-processing includes but are not limited to inverting images, rescaling, binarization,

noise removal, dilation, and erosion, rotation / deskewing, border removal, etc. [44].

To avoid all the ways your tesseract output can drop, you need to make sure the image is

appropriately pre-processed. Given that our data is custom generated, we typically had a clear

and sharp images devoid of noise. As a result, we did not require pre-processing except for

rescaling of the image to our model’s specification.

Figure 13. Image before pre-processing [43].

 41

Figure 14. Image after pre-processing [43].

Regular Expressions

A regular expression (regex or regexp for short) is a special text string for describing a

search pattern. You can think of regular expressions as wildcards on steroids. You are probably

familiar with wildcard notations such as *.txt to find all text files in a file manager. The regex

equivalent is «.∗\. 𝑡𝑥𝑡» [45]. Regular expressions offer a limited but powerful metalanguage to

describe all kinds of formats, protocols, and other small textual languages. Regular expressions

arose in the context of formal language theory, and a primary use has been as part of scanners in

compilers. However, nowadays their applications extend far beyond those areas. For example,

regular expressions have been used in editors for structured text modification [46]. They are used
 42

in network protocol analysis and for specifying events in a distributed system. Regular expres-

sions are used for virus detection using signature scanning, in mining the web, and as alternatives

types for XML data [46]. Moreover, there are many uses of regular expressions outside of com-

puter science, for example, in sociology (for characterizing events that led to placing of a child in

foster care) or biology (for finding DNA sequences) [46]. In addition to specific applications,

many generic types of information, such as phone numbers or dates, are often presented and ex-

changed in specific formats that are described using regular expressions [46].

Regular expressions, though powerful, are often problematic due to its complexity (ob-

scure, and hard to understand and reuse), errors (subtle and hard to detect faults), and version

proliferation (making it hard to select the right one for a specific task). Some complex regular

expression could exceed 100 characters, while some can go as high as over 4000 characters. This

makes it difficult to scale regular expressions.

 43

Chapter III: Experiment

Overview

In real-life application, speed, accuracy, and resources tradeoffs must be made. The inter-

net is a great resource for different classes of trained data ready for use. Unfortunately, there is

not one class ready made for recognition of URLs in an image. Most tests and trainings are done

with datasets, and their results are measured in mean Average Precision (mAP) at Intersection

over Union (IoU) threshold. IoU measures the overlap between two regions.

In this Chapter, we will highlight the custom generation of our training data from our

word bank of 56000 (Figure 15) words.

Data Preparation

A dataset is a collection of data. Image datasets are used to train and benchmark object

detection algorithms [40]. For the algorithm to be trained, we need to create images with URLs

in them and store the coordinates of the URLs. This would be used by the training algorithm to

know what is right (a URL) and use that to learn what is wrong (not a URL).

Python was our choice programing language and because YOLO is just a technique for

detection using convolutional neural network, we made use of pythons’ keras library as well.
 44

Figure 15. Word bank used to generate our training image data.
Figure 16 shows a sample image generated from our word bank. This process is known as

tagging, and ours is automated. A total of sixty thousand images were generated for this experi-

ment, each having varying number of URL(s) in them and some having no URL at all, giving the

algorithm a large enough sample size. A record of the x and y coordinate of any URL generated

is stored and is later passed into our detection algorithm.

Figure 16. Sample generated image using our custom-built tagging tool.
 45

Figure 17. Sample grid generated by YOLO for our detection algorithm.
Training

Training, according to Longman Dictionary [34], is the process of teaching or being

taught the skills for a particular job or activity. In this context, it refers to the process of teaching

an algorithm towards a specific task for which it will be used. These algorithms also learn from

experience without being explicitly programmed. In our experiment, we labelled ten thousand

images, each with varying number of URLs and some with no URL at all. We trained our algo-

rithm using fifty thousand images. These images have varying font size. Our image is sectioned

into 21 × 21 grid cells (Figure 17), each is responsible for predicting K bounding boxes. The

grid cell was selected because we would be working with only text. An object is considered to lie
 46

in a specific cell only if the center co-ordinates of the anchor box lie in that cell. Due to this

property, the center co-ordinates are always calculated relative to the cell, whereas the height and

width are calculated relative to the whole image size. Using the Equation below, YOLO deter-

mines the probability that a cell contains a certain class. The class with the maximum probability

is chosen and assigned to that grid cell. This is repeated for all grid cells in the image. The proba-

bility that there is an object of certain class ‘c’ is:

𝑠𝑐𝑜𝑟𝑒𝑐 ,𝑖 = 𝑝𝑐 × 𝑐𝑖

Setting 0.3 filter size and 0.7 𝐼𝑜𝑈 threshold, we applied non-max suppression to select

the bounding box with the highest 𝐼𝑜𝑈 threshold. Batch normalization is then applied to add

noise to the inputs of every later, discouraging overfitting preventing our model from producing

deterministic values for a given training example. We later applied max pooling to down-sample

out input (the batch normalized output) followed by our activation function. We settled with Rec-

tified Linear Units (ReLU) – see Figure 18.

At the end of the training process, provided the trained algorithm with a new set of data

(with ten thousand images) to test our result. This new dataset is modelled like the training data.
 47

Figure 18. Our detection model summary.

Evaluation Metric

Intersection over Union (IoU) is an evaluation metric (between 0 and 1) that measures the

overlap between the boundaries of the ground truth of an annotation and the predicted boundary.

IoU evaluates whether a prediction is “good enough” [36]. The closer the prediction is to 1, the

closer to perfect it is. Figure 19 illustrates the graphical view of the equation below.

𝐴𝑟𝑒𝑎 𝑜𝑓 𝐼𝑛𝑡𝑒𝑟𝑠𝑒𝑐𝑡𝑖𝑜𝑛/𝑂𝑣𝑒𝑟𝑙𝑎𝑝
𝐼𝑜𝑈 =
𝐴𝑟𝑒𝑎 𝑜𝑓 𝑈𝑛𝑖𝑜𝑛

Figure 19. Graphical View of the IoU equation [41].

 48

In general, when 𝐼𝑜𝑈 ≥ 0.5, we consider the prediction correct. Knowing IoU, we calcu-

late precision and recall. We also calculated the confidence score. Confidence score reflects how

likely the box contains an object and how accurate the bounding box is.

Figure 20. The higher the IoU, the better the performance [24].
 49

Chapter IV: Results

Overview

The training of our object detection algorithm was over 200 epochs. It took approxi-

mately 8 hours, efficiently making use of a core i7 NVIDIA GeForce GTX 1650 with Max-Q

Design computing power. During each epoch, the confidence metric improved over time, making

detection closer to perfect.

Figure 21. (a) Object detection algorithm after 10 epochs (b) Object detection algorithm for an
image without any link (c) Improved Object detection algorithm after 200 epochs.
Detection Accuracy

Using the online Kaggle training compute power, we were able to achieve better accu-

racy. Running the training on 500 epochs brought the result close to perfect.
 50

Detection Accuracy on Regular Font Styles

6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 10
Actual Thesis OCR pattern 1 OCR pattern 2 - regex

Figure 22. Accuracy between two Optical Character Recognition techniques and our model on
ten im-ages using regular font style.
The OCR patter 1 technique used is a python library called “urlextract” while for the

other pattern, we used regular expression. Before selecting the regular expression to use, we tried

different ones from stackoverflow and chose the most accurate (see Figure 23).

Figure 23. Regular Expression used to recognize a URL in a string of text.

The reason for our result in image 1 (see Figure 24) detecting an extra image is because

of overfitting. However, if converted into text, would be quickly eliminated because of the nature

of the area of interest (see Figure 25). With improved tunning and more training, we should ex-

perience better accuracy.

 51

Figure 24. Our model experiencing overfitting.

The observed reason for the errors when the urlextract library and the regular expression

technique were used is, the Optical Character Recognition (OCR) accuracy. The OCR we used is

the pytesseract image to string library. It is a deep learning model that detects, recognizes and

outputs texts based on learned patterns. This technique could get worse in an unconstrained envi-

ronment where the level of noise and distortions is not controlled.

 52

Figure 25. OCR image conversion to text versus our object detection model.
Font Style and Size

Font style and size significantly affects the visibility of a text. This is also an important

factor to consider when creating a computer vision model to recognize texts. There are different

categories of font styles. We will consider serifs and sans-serifs briefly. A serif is a decorative

stroke that finishes off the end of a letters stem (sometimes also called the “feet” of the letters). A

sans serif is a font that does not.

OCR software “won the battle” on italicized texts. This is simply because our model was

trained on strictly serif and sans-serif fonts. Verdana (sans-serif), Arial (sans-serif), Segoe UI

(sans-serif), Calibri (sans-serif), Cambria (serif) font style were used to train our model. The font

size of our training images ranged between 14 pixels to 18 pixels. Our algorithm got better accu-

racy when the test images have similar texts font style.

In Figure 27, our model was evaluated on about 100 images with texts of font family of

Brush Script MT and Bauhaus and font size between 24 and 28.
 53

Figure 26. Example italicized text.

Detection Accuracy on italicized texts

80
70
60
50
40
30
20
10
0
Thesis OCR Pattern 1 OCR Pattern 2 - regEx

Errors Overfitting Exact

Figure 27. Detection Accuracy between Optical Character Recognition techniques and our
model on ten images using italicized font style.
Training our model on these variants of font styles (and font size) should fix the less ac-

curacy gotten from italicized texts in images.

 54

Input Size

Using 400 x 400 input size image for training, we tested our algorithm on larger image

sizes 1200 x 1200 and 600 x 600. Given that pytesseract OCR has been trained on a wide range

of data, our model failed to perform optimally.

Figure 28. Detection on different input image sizes.

For better performance, comprehensive training is required with larger data set consisting

of variety of input sizes and italicized texts with different font family, all of which pytesseract

has.
 55

Chapter V: Conclusion

Overview

The title of this project is “Object Detection and Recognition Using YOLO: Detect and

Recognize URL(s) in an Image Scene”. This translates into the usage of object detection and

recognition technology to successfully identify URL in images.

To achieve this, a custom YOLO algorithm is trained on fifty thousand images and evalu-

ated on ten thousand images. The algorithm employs a 21 x 21 grid. When set up this way, it al-

lows the bounding boxes generated to capture the objects properly with minimal chances of over-

fitting.

To generate training data, a text generator was programmed to generate images which

randomly contain both texts and URLs and records useful information about the location of the

URLs in a text file in JSON format. About fifty thousand images are used to train the model in

about 500 epochs, allowing for a very dependable performance model.

Application

The application of the algorithm is quite restricted, with its main use being the detection

of URLs in an image scene. On its own, the application of the model is not wide and varied.

However, in combination with other modifications, such as the augmentation this algorithm can

prove to be very useful and beneficial. The project has a few applications in various fields. These

applications will be described in subsequent paragraphs.

The model can be particularly useful in the field of tourism. Tourist can take a picture of

a URL and have the information of the website displayed without having to type on their key-
 56

board. Business owners, shop owners, and their customers can advertise and publicize infor-

mation by leaving URLs to the services they offer on their advertisement spaces and users can

pick up the URL(s) by the click of a button. The model can also be used to capture reference

URLs while listening to a presentation at a conference. The uniqueness in of this model is that

valuable time need not be wasted by users and eliminating the possibility of not being able to

completely type out a URL in time in the case of a presentation. Given the ability of smart phone

cameras in recent times, taking a picture of a URL while in transit should result in a “not too

blurry” image which can be passed in as an input to our model and the URL(s) of interest re-

trieved.

The software can also be utilized at auctions that provide URL(s) to read more about an

item. With the utilization of this model, information may be passed by appending URL(s) con-

taining the full description of the items linked to them.

Another situation where the use of this model is in museum and zoos. Less space needs to

be taken up when describing an item. A brief description of an item and a link to the full descrip-

tion would make museums and zoos less congested with placards. If implemented in the form of

an application, could provide real-time data on items.

As opposed to QR codes, this model does not require any special software to encode the

URL. QR codes cannot be handwritten but if our model is extended to handwriting recognition,

it would be able to detect URLs.

 57

Limitations of Work

One of the major limitations of this work is handwriting detection. The data which the al-

gorithm is trained with is restricted entirely to digitally written text. As such, URLs written in

freehand have a significantly lower chance of detection by the model.

Furthermore, the algorithm can only recognize horizontal texts. As a result of this, it is

limited in functionality such that vertical as well as diagonal texts will be impossible to detect by

the algorithm.

Due to the type of data the algorithm has been trained with; it is also impossible for the

algorithm to detect links that span multiple lines. This is because the bounding boxes trained are

not big enough to encompass texts that span multiple lines.

Future work

Implementing an OCR alongside our model would make converting URLs in a picture to

clickable text. This would greatly reduce the need for advanced software like QR codes genera-

tors and QR code detectors. A person would need to simply have a URL displayed and anybody

may freely access it when equipped with our modelled algorithm.

The second improvement would be to attempt to increase the range of the URLs that can

be detected by the model. So far, our model was trained on slightly varying font size (see Figure

29). In general, font type should not create a significant difference in the accuracy of our model.

It would be great to detect vertically aligned URLs (although that is not a common case in real

life), and to detect URLs at different angles. Another case would be detecting URLs written

backwards.
 58

Figure 29. Our models performance on slightly varying font size.

Finally, the future iterations of this model could be trained on handwritings and more

noisy environments and fonts to improve its ability to recognize a wider range of texts effec-

tively.
 59

References

[1] P. Chakravorty, "What Is a Signal? [Lecture Notes]," in IEEE Signal Processing Magazine,
vol. 35, no. 5, pp. 175-177, Sept. 2018, doi: 10.1109/MSP.2018.2832195.
[2] M. Rouse, “image” [Online]. Available at: https://whatis.techtarget.com/definition/image.
[Accessed: 01/10/20]
[3] Merriam-Webster “image” [Online]. Available at: https://www.merriam-webster.com/dic-
tionary/image, [Assessed: 01/10/20]
[4] Wikipedia, “Image” [Online]. Available at: https://en.wikipedia.org/wiki/Image, [Assessed:
01/10/20]
[5] The Editors of Encyclopedia Britannica, “Image-processing” [Online], Available at:
https://www.britannica.com/technology/image-processing. [Assessed: 01/10/20]
[6] Wikipedia, “Encoding images” [Online], Available at:
https://www.bbc.co.uk/bitesize/guides/zqyrq6f/revision/3. [Assessed: 02/10/20]
[7] Object (image processing) [Online], Available at: https://en.wikipedia.org/wiki/Object_(im-
age_processing) , [Assessed: 02/10/20]
[8] P. Ganesh, Object Detection: Simplified [Online], Available at: https://towardsdatasci-
ence.com/object-detection-simplified-e07aa3830954, [Assessed: 02/10/20]
[9] Tensorflow, Available at: https://www.tensorflow.org/lite/models/object_detection/overview,
[Assessed: 03/10/20]
[10] Wikipedia, “Object detection” Available at: ‘https://en.wikipedia.org/wiki/Object_detection,
[Assessed: 02/10/20]
[11] Fritz, “Object Detection Guide”, Available at: https://www.fritz.ai/object-detection/, [As-
sessed: 03/10/20]
[12] Wikipedia, “Outline of object recognition”, Available at: https://en.wikipedia.org/wiki/Out-
line_of_object_recognition , [Assessed: 02/10/20]
[13] Object Recognition, Available at: cse.usf.edu/~r1k/MachineVisionBook/Machine-
Vision.files/MachineVision_Chapter15.pdf, [Assessed: 03/10/20]
[14] N. Pinto, D. D. Cox, and J.J. DiCarlo, “Why is Real-World Visual Object Recognition
Hard?” (2008) PLoS Comput Biol 4(1): e27.
[15] A. Gulli and P. Sujit, “Deep Learning with Keras” (2017), Available at:
https://1lib.us/book/3411804/7ea47a?id=3411804. [Assessed: 03/10/20]
[16] “The Definitive Glossary of Higher Mathematical Jargon - Algorithm". Available at:
https://mathvault.ca/math-glossary/ [Assessed: 03/10/20]
[17] “Definition of ALGORITHM". Merriam-Webster Online Dictionary. Available at:
https://www.merriam-webster.com/dictionary/algorithm [Assessed: 04/10/20]
[18] Y. Gavrilova, “Artificial Intelligence vs. Machine Learning vs. Deep Learning: Essentials”
Available at https://serokell.io/blog/ai-ml-dl-difference [Assessed: 04/10/20]
[19] J. Brownlee, “A Gentle Introduction to Object Recognition with Deep Learning” (2018)
Available at: https://machinelearningmastery.com/object-recognition-with-deep-learn-
ing/ [Assessed: 01/10/20]
 60

[20] A. Kamal, “YOLO, YOLOv2 and YOLOv3: All You want to know”(2019) Available at:
https://medium.com/@amrokamal_47691/yolo-yolov2-and-yolov3-all-you-want-to-
know-7e3e92dc4899 [Assessed: 04/10/20]
[21] You Only Look Once: Unified, Real-Time Object Detection, 2015. Available at:
https://arxiv.org/abs/1506.02640, [Assessed: 04/10/20]
[22] A. Rosebrock., “Intersection over Union (IoU) for object detection” (2016) Available at:
https://www.pyimagesearch.com/2016/11/07/intersection-over-union-iou-for-object-de-
tection/, [Assessed: 04/10/20]
[23] I. Tan, “Measuring Labelling Quality with IOU and F1 Score”, Available at: https://me-
dium.com/supahands-techblog/measuring-labelling-quality-with-iou-and-f1-score-
1717e29e492f , [Assessed: 01/10/20]
[24] StackOverflow, “Intersection Over Union (IoU) ground truth in YOLO”, Available at:
https://stackoverflow.com/questions/61758075/intersection-over-union-iou-ground-truth-
in-yolo, [Assessed: 02/10/20]
[25] J. Redmon & A. Farhadi, (University of Washington), “YOLO9000: Better, Faster,
Stronger” Available at: https://pjreddie.com/media/files/papers/YOLO9000.pdf, [As-
sessed: 02/10/20]
[26] “YOLO v2 – Object Detection” Available at: https://www.geeksforgeeks.org/yolo-v2-ob-
ject-detection/ [Assessed: 04/10/20]
[27] A. Aggarwal, “YOLO Explained” Available at: https://medium.com/analytics-vidhya/yolo-
explained-5b6f4564f31 [Assessed: 04/10/20]
[28] K. Mahesh Babu, M.V. Raghunadh, Vehicle number plate detection and recognition using
bounding box method, May 2016, pp 106–110
[29] L. Cai, F. Jiang, W. Zhou, and K. Li, Design and Application of An Attractiveness Index for
Urban Hotspots Based on GPS Trajectory Data, (Fellow, IEEE), pg 4
[30] Wikipedia, “Minimum bounding box” Available at: https://en.wikipedia.org/wiki/Mini-
mum_bounding_box, [Assessed: 04/10/20]
[31] Dive into Deep Learning, “13.3. Object Detection and Bounding Boxes” Available at:
https://d2l.ai/chapter_computer-vision/bounding-box.html , [Assessed: 04/10/20]
[32] J. Redmon & A. Farhadi, YOLOv3: An Incremental Improvement, University of Washing-
ton Available at: https://arxiv.org/abs/1804.02767 [Assessed: 05/10/20]
[33] YOLO: You Only Look Once, Available at: jeremyjordan.me/object-detection-one-
stage/#yolo, [Assessed: 05/10/20]
[34] Longman Dictionary, Definition of training, Available at: https://www.ldoceonline.com/dic-
tionary/training [Assessed: 05/10/20]
[35] Guangrui Liu “Real-Time Object Detection for Autonomous Driving Based on Deep Learn-
ing” (2017), Available at: https://tamucc-ir.tdl.org/handle/1969.6/5637 [Assessed:
03/18/21]
[36] A. Abdulkader & C. Vlahija “Real-time vehicle and pedestrian detection, a data-driven rec-
ommendation focusing on safety as a perception to autonomous vehicles”, Available at:
http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A1479957&dswid=-3676 [As-
sessed: 03/18/21]
 61

[37] Enhancement methods in image processing Available at: https://www.mathworks.com/dis-

covery/image-enhancement.html. [Assessed: 03/18/21]
[38] Evaluating a machine learning model. Available at: https://www.jeremyjordan.me/evaluat-
ing-a-machine-learning-model/. [Assessed: 03/18/21]
[39] A Comprehensive Guide to Convolutional Neural Networks — the ELI5 way. Available at
https://towardsdatascience.com/a-comprehensive-guide-to-convolutional-neural-net-
works-the-eli5-way-3bd2b1164a53. [Assessed: 03/18/21]
[40] J. Jokela “Person Counter Using Real-Time Object Detection and a Small Neural Network”
Turku University of Applied Sciences. Available at: https://www.theseus.fi/bit-
stream/handle/10024/153489/Jokela_Jussi.pdf?sequence=1&isAllowed=y. [Assessed:
03/18/21]
[41] Manishgupta “YOLO – You Only Look Once”. Available at: https://towardsdatasci-
ence.com/yolo-you-only-look-once-3dbdbb608ec4. [Assessed: 03/18/21]
[42] E. Y. Li “Dive Really Deep into YOLO v3: A Beginner’s Guide”. Available at: https://to-
wardsdatascience.com/dive-really-deep-into-yolo-v3-a-beginners-guide-9e3d2666280e.
[Assessed: 03/18/21]
[43] F. Zelic & A. Sable “A comprehensive guide to OCR with Tesseract, OpenCV and Py-
thon.”. Available at: https://nanonets.com/blog/ocr-with-tesseract/. [Assessed: 03/18/21]
[44] Tessdoc. Available at: https://github.com/tesseract-ocr/tessdoc/blob/master/ImproveQual-
ity.md. [Assessed: 06/04/21]
[45] J. Goyvaerts “Regular Expressions: The Complete Tutorial”. Available at: https://www.reg-
ular-expressions.info/print.html. [Assessed: 06/04/21]
[46] M. Erwig & R. Gopinath, “Explanations for Regular Expressions”. Available at:
https://web.engr.oregonstate.edu/~erwig/papers/ExplRegExp_FASE12.pdf. [Assessed:
06/04/21]