Amanuel Worndimu 2015

Addis Ababa University
College of Natural Sciences

School of Information Sciences
A Generalized Approach to Amharic

Braille Recognition
By:
Amanuel Wendmu
Advisor:
Dereje Teferi (Ph.D.)
June, 2015
School of Information Sciences
A GENERALIZED APPROACH TO AMHARIC

BRAILLE RECOGNITION
BY: AMANUEL WENDMU
ADVISOR: DEREJE TEFERI (PHD.)
A Thesis Submitted to the School of Graduate Studies of the Addis

Ababa University in Partial Fulfillment for the Degree of Master of
Science in Information Science
June, 2015
School of Information Science
A GENERALIZED APPROACH TO AMHARIC

BRAILLE RECOGNITION
BY: AMANUEL WENDMU
Name and Signature of the members of examining board

Name Title Signature Date
Dr. Dereje Teferi Advisor
Dr. Martha Yefru Examiner
Dr. Million Meshesha Examiner
Dedicated to my family
i
ACKNOWLEDGEMENT
First and for most, I would like to thank God, the almighty for making this possible.
I am also thankful to my advisor Dr. Dereje Teferi for his guidance and support in the course
of this work.
I also like to thank my friend Berhanu Sahle for the invaluable ideas and comments he
provided.
I am also heartily thankful to employees of Misrach Center Ato Geremew and Ato Amare
who teach me about braille writing, give their time unreservedly to respond to my questions
and provided me the Amharic brialle documents that I have used as a dataset in this work.
My greatest gratitude goes to my loving parents Wendmu Zengu and Letensea Zemichael
who brought me to this world, raised and educated me. I also want to thank my brothers
Nigusse Wendmu and Fanuel Wendmu and sisters Misgana Wendmu and Milite Wendmu
who loved me, shared me their ideas and supported me all the way.
I also would like to thank my fiancé Helina Getachew whose love, care and encouragement
have carried me through my difficult times.
Finally, I would like to thank Miftah Hasssen for all the resources he shared and my friends
Binyam Gedle, Yonas Tadele and Gedion Assefa who were providing me resources and ideas
in the course of my years at AAU.
Amanuel Wendmu
ii
Table of Contents
List of Figures ................................................................................................................................... vi
List of Tables ....................................................................................................................................vii
List of Algorithms ............................................................................................................................ viii
List of Abbreviation ...........................................................................................................................ix
ABSTRACT ......................................................................................................................................... x
CHAPTER ONE ................................................................................................................................... 1
INTRODUCTION ................................................................................................................................. 1
1.1. Background ........................................................................................................................ 1
1.1.1. Braille Writing System ................................................................................................ 1
1.1.2. Braille Recognition Systems ........................................................................................ 3
1.2. Statement of the Problem .................................................................................................. 5
1.3. Objective of the Study ........................................................................................................ 8
1.3.1. General Objective ....................................................................................................... 8
1.3.2. Specific Objectives ...................................................................................................... 8
1.4. Scope and Limitation of the Study ...................................................................................... 8
1.5. Methodology of the Study .................................................................................................. 9
1.5.1. Literature Review ....................................................................................................... 9
1.5.2. Dataset Collection .................................................................................................... 10
1.5.3. Implementation tool................................................................................................. 10
1.5.4. Testing Procedure .................................................................................................... 11
1.5.5. Tools Used ......................................................................................................................... 12
1.6. Significance of the Study .................................................................................................. 13
1.7. Organization of the study ................................................................................................. 14
CHAPTER TWO ................................................................................................................................ 15
LITERATURE REVIEW........................................................................................................................ 15
2.1. History of Braille............................................................................................................... 15
iii
2.2. Braille Reading and Writing .............................................................................................. 16
2.3. Double-Sided Braille ......................................................................................................... 19
2.4. Amharic Writing and Braille System.................................................................................. 20
2.4.1 Amharic Writing System ........................................................................................... 20
2.4.2 Amharic Braille ......................................................................................................... 24
2.5. Automatic Braille Recognition .......................................................................................... 27
2.5.1 Image Acquisition ..................................................................................................... 27
2.5.2 Preprocessing ........................................................................................................... 27
2.5.3 Braille Dot Detection ................................................................................................ 28
2.5.4 Braille Cell Formation ............................................................................................... 29
2.5.5 Interpretation........................................................................................................... 29
2.6. Review of Braille Recognition Works ................................................................................ 29
2.6.1 Review of Amharic Braille Recognition works ........................................................... 29
2.6.2 Review of double-sided Braille recognition works ..................................................... 32
CHAPTER THREE .............................................................................................................................. 37
AMHARIC BRAILLE RECOGNITION .................................................................................................... 37
3.1. General Design of Amharic Braille Recognition ...................................................................... 38
3.2. Image Acquisition/Digitization .............................................................................................. 40
3.3. Preprocessing of double-sided Braille Images........................................................................ 40
3.4. Segmentation ....................................................................................................................... 41
3.4.1. Braille Dot Detection ...................................................................................................... 42
3.4.2. Separation of recto and verso ........................................................................................ 42
3.4.3. Braille Cell Grouping ....................................................................................................... 44
3.5. Feature extraction ................................................................................................................ 47
3.6. Translation............................................................................................................................ 48
CHAPTER FOUR ............................................................................................................................... 50
EXPERIMENTATION ......................................................................................................................... 50
iv
4.1. Data collection ...................................................................................................................... 50
4.2. Digitization / Image Acquisition............................................................................................. 51
4.3. Thresholding ......................................................................................................................... 53
4.4. Separation of dots for Double-Sided Braille ........................................................................... 55
4.5. Braille Cell formulation ......................................................................................................... 58
4.6. Amharic Braille feature extraction......................................................................................... 61
4.7. Translation............................................................................................................................ 63
4.8. Performance Evaluation ........................................................................................................ 65
4.9. Discussion and Challenges..................................................................................................... 69
CHAPTER FIVE.................................................................................................................................. 71
CONCLUSION AND RECOMMENDATION .......................................................................................... 71
5.1. Conclusion ............................................................................................................................ 71
5.2. Recommendation ................................................................................................................. 73
Reference ........................................................................................................................................ 74
Appendix ......................................................................................................................................... 78
I. List of the seven braille orders ............................................................................................. 78
II. Complete list of Amharic Alphabet ....................................................................................... 79
III. First Version Amharic Braille ............................................................................................ 80
IV. Second Version Amharic Braille ........................................................................................ 81
V. Third Version Amharic Braille ............................................................................................... 82
VI. Fourth Version Amharic Braille ......................................................................................... 83
I. Lookup Table ....................................................................................................................... 84
II. Sample Amharic braille and Translation ............................................................................... 86
v
List of Figures
Figure 1.1 Six dot braille positioning of recto and verso ....................................................... 2
Figure 1.2 Standard distance in braille document .................................................................. 2
Figure 1.3 Braille Recognition Process ................................................................................. 4
Figure 2.1 English Braille alphabets, numeral and punctuation marks ................................. 18
Figure 2.2 Part of the Amharic Script ................................................................................. 23
Figure 2.3 Part of Amharic Alphabet Braille representation ................................................ 25
Figure 2.4: Fourth Version Amharic Braille Numerals representation .................................. 26
Figure 2.5 Fourth version Amharic Braille Punctuation signs .............................................. 26
Figure 3.1 images of recto and verso dots ............................................................................ 42
Figure 4.1 Scanned braille image; ....................................................................................... 52
Figure 4.2 MATLAB code for converting full color image into gray scale .......................... 53
Figure 4.3 MATLAB code for thresholding ......................................................................... 54
Figure 4.4 Sample thresholded image .................................................................................. 55
Figure 4.5 Part of MATLAB code for front and back side dot separation ............................ 56
Figure 4.6 Recto and Verso dots separation ......................................................................... 57
Figure 4.7 Binary braille dote regions before and after morphological improvement ........... 58
Figure 4.8 Part of the MATLAB code for building grid ....................................................... 59
Figure 4.9 Results of grid construction ................................................................................ 60
Figure 4.10 Results of grid refinement................................................................................. 61
Figure 4.11 MATLAB code for feature extraction ............................................................... 62
Figure 4.12 Steps of translation ........................................................................................... 64
Figure 4.13 Segmentation results ......................................................................................... 67
Figure 4.14 Problems identified in thresholding .................................................................. 69
vi
List of Tables
Table 1.1 confusion matrix .............................................................................................................. 12

Table 2.1 Fourth version Amharic Braille Variants ............................................................................ 25
Table 4.1 Database created ............................................................................................................. 51
Table 4.2 Sample Amharic features and their normalized representation ........................................ 63
Table 4.3 Performance results for segmentation single-sided .......................................................... 65
Table 4.4 Performance results for segmentation double-sided ........................................................ 66
Table 4.5 Overall performance achieved single sided ....................................................................... 68
Table 4.6 Overall performance achieved double-sided..................................................................... 68
Table I List of the seven braille orders .............................................................................................. 78
Table II.1 Complete list of Amharic Alphabet ................................................................................... 79
TableIII.1 List of basic Amharic Braille characters in the first version ................................................ 80
TableIII.2 List of vowel variant for Amharic Braille characters in the first version ............................. 80
Table IV.1 List of basic Amharic Braille characters in the 2nd version ............................................... 81
Table IV.2 List of vowel variants for Amharic Braille characters in the 2nd version ........................... 81
Table V.1 List of basic Amharic Braille characters in the 3nd version ................................................ 82
Table V.2 List of vowel variants for Amharic Braille characters in the 3nd version ............................ 82
Table VI.1 List of fourth Version Amharic Braille punctuation marks ................................................ 83
vii
List of Algorithms
Algorithm 3.1 Thresholding algorithm ............................................................................... 41

Algorithm 3.2 Front and back side dots separation algorithm ............................................. 43
Algorithm 3.3 Grid construction algorithm .......................................................................... 45
Algorithm 3.4 Grid refining algorithm ................................................................................. 46
Algorithm 3.5 feature extraction algorithm .......................................................................... 47
viii
List of Abbreviation
ABR Automatic Braille Recognition

AAU Addis Ababa Univeristy
ANN Arteficial Nueral Network
BMP Bit MaP
DPI Dots Per Inch
EABS Ethiopian Association for Blind Society
HP Hawlet Packard
MATLAB MATrix LABoratroty
OBR Optical Braille Recognition
OCR Optical Character Recognition
RGB Red Green Blue
SVM Support Vector Machine
UNESCO United Nations Educational, Scientific and Cultural Organization
WBA World Blind Association
WHO World Health Organization
ix
ABSTRACT
In 1825 Louis Braille invented the writing system braille which is the preferred means of
written communication for blind people around the world. In Ethiopia, there has been a large
production of Amharic braille documents from the time of its introduction in 1934. Since
Braille documents are usually read and understood only by visually impaired and those others
who have learned braille only, their literatures are highly restricted from reaching the sighted
society. Optical Braille Recognition (OBR) is an important technology to bridges the
communication gap that exists between the blind and the sighted people. In addition, OBR
serves as a technology to preserve documents that only exist in Braille form and can also
serve as a simple mechanism for braille copy.
Past researches carried out locally have made an attempt to develop an ABR system for
Amharic braille recognition. There were also some efforts made to improve the performance
of the Amharic ABR system. As a continuation to those previous efforts, this research
presented an approach for recognition of double-sided Amharic braille documents. In this
regard, image thresholding and segmentation techniques are explored and adopted in this
work with some improvements.
An adaptive thresholding algorithm that uses the pixel values on every image is used to
compute two threshold values that segment the Braille image into bright, dark and
background. A moving window scanning technique that scan every column of an image and
extract the dot regions for the front and back side of the image and put them into two separate
arrays is also adopted and improved. In addition, an improved grid construction technique is
designed that automatically determines the average dot area and exclude dots which are
smaller or bigger in size that can affect the performance of the grid construction. The grid
construction process has incorporated techniques for refining itself when any error exists.
Braille cells are formulated by using the location of six intersection points in the constructed
grid. Finally, a lookup table is developed for every character in Amharic and features of
braille cells are matched against corresponding Amharic print characters.
The system is tested using database from scanned clean, average noisy and high level noise
Amharic braille documents. The developed system works with 100% accuracy for both
single-sided and double-sided clean braille documents. However, as the level of noise
increases the performance of that system fails. Overall, an accuracy level of 99.27% for
x
single-sided and 88.16% for double-sided braille documents is achieved. Degradation of
braille document and skewness has been the major challenges. The results achieved show that
the work is promising. However, to further improve the system, areas that need further
consideration in future such as improved image filtering, skew angle correction, format
preservation and post recognition correction are recommended.
xi
CHAPTER ONE
INTRODUCTION
1.1. Background
According to WHO‟s 2014 statistics, over a billion people, about 15% of world‟s population,
have some form of disability; 285 million people are estimated to be visually impaired
worldwide of which 39 million people are blind and 246 million people have low vision.
About 90% of the world‟s visually impaired and blind people live in low income setting [49].
Visually impaired people are part of the society that plays a significant role where they live.
Therefore, it is a must to provide this people with a means and system that enable them to
communicate with the world. There are different mechanisms created for visually impaired
people as a means to provide them what the world has in printed document. One of the most
valuable and indispensable method is through the use of braille [3].
1.1.1. Braille Writing System
Braille has been the preferred means of writing by visually impaired people since its
introduction in 1825 by the French scientist Louis Braille. The innovation of braille has
created an efficient means of communication for the blind and visually impaired people and
allowed them access to education and employment [12].
Braille consists of a combination of punched six dots placed in a fixed two by three matrix
called cells. Six dot writing is widely used giving a total of 64 (26) possible combinations of
raised dots in a single cell that can be assigned for different characters or words. Braille has a
very low information density that an average page of 26 x 34 cm sheet can only have 27 lines
and 32 characters per line. Like the ordinary writing used by sighted people, it is also possible
to write on both sides of a braille paper by shifting the text on the verso so that it will not
conflict with the recto [26].
Figure 1.1 shows how dots on the verso are placed between the recto dots.
1
Figure 1.1 Six dot braille positioning of recto and verso [26]
People have different annotations for braille characters that vary between language and
domain. But, the space (distance) between points, between characters and between lines is
standard as shown in figure1.2 below [3] [4]. The dimension of point in accordance with the
braille dimension and tactile resolution of the finger tips of people has been defined and
specified for worldwide use. Horizontal and vertical distance between points in the letter, in
addition to, the distance between the cells that represent a word after completion are specified
by the library of US Congress [23].
 The horizontal distance (Ph) and vertical distance (Pv) between dots within a cell is
2.5 millimeter,
 The horizontal distance or character space (Rh) between cells is approximately 3.5
millimeter,
 The vertical distance or line space between cells is approximately 5 millimeter.
Figure 1.2 Standard distance in braille document [16]
2
Though the American and France braille styles are the two main types of braille styles, most
countries have adopted and/or defined their braille code so as to fit their local language
characters. In general, braille systems are categorized into two levels, Grade 1 and Grade 2
[42].
In the Grade 1 braille representation, each possible arrangement of dots within a cell
represents only one letter, number, punctuation sign, or special braille composition sign. This
is the basic representation of letters where each letter is to be represented by one cell
containing two columns where each column contains three rows. In Grade 1 braille,
individual cells cannot represent words or abbreviations which make documents produced in
this grade of braille to be bulkier [42].
Grade 2 braille was introduced as space-saving alternative to Grade 1 braille. Since braille
books are much larger than normal text books a contraction is used to reduce the length of
some words. Therefore, in Grade 2 braille, a cell can represent a shortened form of a word.
Many cell combinations have been created to represent common words, such as brl for
braille, ab for about, alr for already and many others. In addition to reducing the size of
books, Grade 2 braille permits faster reading and this makes Grade 2 braille to be preferred
over Grade 1[11].
The nature of braille writing has made braille documents to be bulkier. One attempt made to
make braille more concise was the introduction of braille contraction; while the other being
the introduction of double-sided braille writing. Double-Sided braille writing can help reduce
the size of bound braille documents such as books. But, it has an impact on braille
recognition performance.
In a given page of inter-point braille there exist protrusions (dots of the recto) and
depressions (dots of the verso). In the scanned images of braille, the dots of protrusions
appear as a black region from above and a brighter region from below while the depressions
appear in reverse order. The major challenge unique to double-sided braille recognition is that
regions of the same type; for instance, white region form protrusion and white region from
depression, are merged creating confusion in the detection process [4] [50].
1.1.2. Braille Recognition Systems

Over the period, there are many researches being conducted in order to exploit the power of
information technology to improve the livings of the society [43]. Accordingly, Optical
3
Character Recognition (OCR) is a system designed to convert scanned printed or handwritten
text image file into readable and editable text document. OCR system receives file as an
image and convert it by matching or comparing features against set of characters stored in
OCR database [23] [28].
OCR has opened a new light of life for blind people. It can be used to allow text-to-braille
and braille-to-text conversion. But, due to the fragmented nature of the characters and the fact
that characters of both sides can be seen simultaneously, braille translation in to text cannot
be processed with standard OCR software. Therefore, a little different approach, OBR, was
devised. OBR is a variation of OCR and is used to digitize and reproduce texts that have been
produced with non-computerized systems, such as braille type writers. OBR exploits the
fixed matrix positioning property of braille dots [23] [27].
As can be seen in figure 1.3 below, braille recognition system has similar processes as OCR
and according to [13] it consists of six operations; Image acquisition, Pre-processing,
Segmentation, Feature extraction and Interpretation.
Figure 1.3 Braille Recognition Process [13]
The first phase, scene constraint, has the aim of maximizing the use of prior knowledge of the
scene and lowering the problem of image analysis in any way possible. A scene constraint
exploits and imposes the environment constraints aiming to reduce the complexity of all
subsequent operations to manageable level. This is achieved by maximizing the use of prior
knowledge of the scene by exploiting existing knowledge and minimizing the problem of
image analysis by effective imposition of constraints [13].
4
Image acquisition is the process of inputting the braille text to our computers as an image
through the use of digital camera or scanner. This process is similar to the way we input
traditional text file by using a scanner, a digital camera or some other input device. There are
different scanner standards available in the market that makes the digitization process easy.
However, the quality of images received may not be appropriate to start processing of these
images [4] [13].
The images we get from scanners are not perfect and free of noise. Therefore, before any
further braille processing for recognition, preprocessing operations such as noise filtering and
edge enhancement, are performed on the images to improve the quality of the image and
prepare the pixel values for subsequent operations so that the segmentation, feature extraction
and classification process that will follow will be simplified [2] [4].
Segmentation separates the preprocessed image in to foreground (dots) and background

regions. There are different segmentation techniques that are particularly applied for braille
image. Mesh-grid and local measure are the two most common. Mesh grid widely makes use
of the distance rule in braille writing to construct a vertical and horizontal grid on strictly
arranged braille Cell [13].Since dots produce different color intensity at top and bottom, local
measure works by making use of this difference in intensity level to define a threshold for
segmentation [3] [41].
Feature extraction process extracts the braille dots from the segmented image and groups
them into cells. There are different feature extraction algorithms. Some algorithms work on
the contours of the image while other algorithms calculate every pixel of the image. But, a
good feature extraction algorithm is one that can produce an image feature that is robust to
transformation and noise as well as efficient in extracting and matching features [13] [40].
The last phase, interpretation, converts the braille cells into their corresponding text. For each
dot, the distance between the centeroid and possible neighbors are determined and from this
information the dots are grouped into cells. The dot patterns in each cell are represented by a
bit string and interpretation is made by searching a corresponding value in the braille
dictionary [13].
1.2. Statement of the Problem
In Ethiopia, there are an estimated half a million visually impaired people that include
students, teachers, artists, lawyers and employees that have significant contribution in
5
political, religious, economic, social and many other affairs of the society. Braille is the
means of codifying knowledge for these people. But, only a small number of sighted people
understand braille. Therefore, the works of visually impaired people is accessed only by those
who understand braille. This has created a wide generation gap between the visually impaired
and vision society [18] [43].
There is massive collection of Amharic braille documents that have been produced since
Amharic braille is introduced in 1934 [52]. These documents are found at different places
like Addis Ababa University Kennedy library Braille Documentation, Misrach Center, Sebeta
visually impaired school, German Church visually impaired school, Ethiopian Association for
Blind Society (EABS), Entoto Blind People School and in different regions of the country.
Digitizing of these documents not only facilitates the communication with visually impaired
but also help preservation of historic braille documents [18] [41] [43].
Literature reveals that there are a lot of attempts made to develop Braille Character
Recognizer for different foreign languages, including English [4] [11], Arabic [2] [16] [24],
Spanish [50] and Chinese [30]. As a result of these efforts, some countries are able to build a
well developed braille recognition system for their local languages. However, none of the
available recognition systems developed so far in foreign countries support Amharic braille.
Therefore, it is a must to have OCR system for local languages like Amharic that addresses
the unique nature of the language.
The attempt for applying OCR technology in recognition of Amharic braille documents was
started in 2009. So far, few researches have been carried out locally towards developing OBR
for Amharic. To the best of the researchers‟ knowledge the work of Teshome [43] was the
first attempt in developing Amharic Braille Recognizer. Teshome [43] developed and tested
his system on a clean braille document, which is not the case in most real life braille
documents due to its tactile nature. As a result, Ebrahim [18] designed a system that explores
different noise detection and removal algorithms in order to enhance the performance of the
system on real life documents. So, he has used Gaussian filtering with adaptive histogram
equalization and morphological operation to detect and remove noise. The effort by Shumet
[41] was to enhance the Amharic Braille Recognition system by focusing on feature
extraction and classification. With these regard, three feature extraction algorithms: fixed cell
measure, horizontal and vertical projection and grid construction were experimented. The test
result shows that the fixed cell measure performs very well. Miftah [32] presented an
6
approach for recognition of Amharic documents using direction field tensor that uses
Gaussian filters and derivative of Gaussian for noise removal and isolation of braille dots
from background. He has also designed a skewness correction technique, which exploits the
horizontal direction nature of braille dots. The most recent work by Berihu[9] was to improve
the Amharic braille recognizer and extend it to work for Tigrigna braille recognition as well.
All attempts made towards recognition of Amharic braille so far are on recognition of single-
sided braille document. This is because double-sided braille writing was not commonly
applied in Amharic braille. But, as Ebrahim [18] mentioned, with the advent of Amharic
braille printer, double-sided braille writing is widely being used. Therefore, in order to have a
fully function Amharic Braille Recognition System, previous researches recommended
designing a system considering double sided documents. As a result, this study has attempted
the recognition of double-sided Amharic braille documents which, to the best of the
researcher‟s knowledge, have not been studied so far.
In effort to solve the stated problem, this study has addressed the following research
questions:-
 What are the special features of double-sided Amharic braille documents?
 What kind of preprocessing techniques are required to prepare the dataset for
recognition?
 What separation technique is suitable to achieve higher performance in separating

front and back side dots in double-sided writing?
7
1.3. Objective of the Study
The general and specific objectives of this study are described here under.
1.3.1. General Objective

The general objective of this project is to explore techniques and algorithms for developing
double-sided Amharic braille recognition system that can enable to convert the content of
braille into textual format.
1.3.2. Specific Objectives

To meet the general objective stated above the following specific objectives are set and
addressed:
 To review previous efforts made towards the development of OCR for Amharic
braille document.
 To collect relevant Amharic braille document and prepare the dataset required for
training and testing the classifier.
 To identify special features of double-sided braille characters in character image

recognition.
 To explore algorithm for recognition of double-sided Amharic braille.
 To develop a prototype for optical braille recognition for Amharic braille.
 To evaluate the performance of the prototype developed.
 Provide conclusion and the way forward for future works.
1.4. Scope and Limitation of the Study
This study is a continuation of previous efforts in the area of Amharic braille recognition.
Past researches carried out locally by Teshome [43], Ebrahim [18], Shumet [41] Miftah[32]
and Berihu[9]; all attempted recognition of single-sided Amharic braille. The scope of this
study limited to exploring techniques suitable for recognition of both single-sided and
double-sided Amharic braille documents which have not been attempted so far. To this end,
various double-sided braille recognition techniques are explored and the best ones are
selected and integrated to the Amharic OBR system.
8
Amharic braille characters can be represented with either one, two or three characters. In this
work we have included all representations and the system is tested on all characters, numerals
and punctuation marks in the Amharic writing system. There was no organized document
image corpus that can be used in this work. Therefore, to test the developed system, a
database that contains both single sided and double sided Amharic Braille document is
created. The created database contains braille images classified in to three noise level; clean,
medium noisy and high noisy. But, due to limitation in time the test is made only on 30
double-sided and 6 single-sided documents only.
Braille can be written on paper sheet or plastic. Double-sided Braille are prepared only with
paper. The embossing process can be done with either a typewriter or a braille embosser. Due
to its availability, only braille documents prepared with the help of embosser are used in this
work.
1.5. Methodology of the Study
Methodology provides a way to achieve the objectives of a research problem. The type of
research design that is used in this study is experimental design. Experimental, also called
empirical research is a type of research design that relies on experience or observation alone,
than any abstract idea or theory. It is a data based research that requires setting up hypothesis
and coming up with facts to prove or disprove the hypothesis and conclusions which are
capable of being verified by observation or experiment [10]. The nature of Experimental
research design used in this work is mixed qualitative and quantitative. it is qualitative in the
sense that is exploratory in nature and provides a complete detailed description of the
research topic but it is also quantitative since it focuses on counting and classifying features
and constructing models and figures to explain what is observed.
In order to undertake and achieve the objective of this research the following methods were
followed.
1.5.1. Literature Review

Literature review is conducted on published information and data relevant to a research
question to summarize the state of the art in the field. It provides an up-to-date account and
discussion of the researches carried out in a particular topic. It helps to test our research
question against what is already known about the subject. Therefore, previous views
expressed by Authors in OCR and OBR in general and Amharic Braille Recognition in
9
particular is collected from published sources such as books, thesis, journal articles,
proceeding and the internet and were reviewed and discussed.
1.5.2. Dataset Collection

There is a large collection of Amharic braille document in AAU-Kennedy Library, Misrach
Center and Entoto Blind School. These centers have served past researchers in Amharic
Braille Recognition by providing Amharic braille Documents to be used as data set.
For the purpose of this research, 30 double-sided and 6 single-sided Amharic Braille
documents are collected from Misrach Center and are used as a dataset to experiment the
developed system. The collected braille documents are scanned using HP flatbed scanner
with a resolution of 200 dpi.
The focus of this research is the recognition of double-sided Amharic braille. But, single-
sided braille documents were also collected. This is because it is believed that the system to
be developed should be tested and proved to work on a single-sided braille documents as
well.
1.5.3. Implementation tool

As discussed above, an HP flatbed scanner is used to create images of braille from the
collected braille document. These images are not free from noise and error. Therefore,
preprocessing activities such as image dimensionality reduction, noise filtering and intensity
reduction, that enable cleaning noise from images of double-sided braille document are
applied. The preprocessing activity is performed by integrating new preprocessing techniques
with algorithms developed by past researches using MATLAB image processing toolbox.
MATLAB is selected because it is easy to use and have extensive library packages and
commands for image processing.
In addition to the preprocessing activity, the segmentation operation, that uses thresholding
and local analysis to segment the dots from background, and feature extraction processes, that
uses horizontal and vertical projection technique with grid construction are also performed on
MATLAB. However, for the simplicity of operation with Unicode characters Python is used
to match features extracted against their equivalent Amharic textual representation.
10
1.5.4. Testing Procedure
The collected Amharic braille documents are classified into three dataset. Every dataset
contains images of twelve Amharic Braille documents, 10 double sided and 2 single sided.
The first dataset contains images of clean Amharic braille document while the second
contains images of medium noisy Amharic braille documents while the third contains high
level noisy Amharic braille documents.
The system was experimented and its performance was tested at two stages. The first test was
done after completion of segmentation process to measure the number of correctly segmented
dots against number of dots in the original text. The other test was performed to measure the
overall performance of the system. At this level we used character level matching to check
the number of correctly translated characters against the total number of characters on the
original braille page.
After manually counting the results, an accuracy measure that shows the percentage of
correctly detected instances normalized against the expected number of result is used.
There are no standard measures that are good for evaluating OBR system [22]. Various
researches have used accuracy to measures the performance of their OBR system. As shown
in formula3.2 below accuracy is the percentage of correct instances identified by the system
against the expected number of instance [45].
𝑖𝑛𝑠𝑡𝑎𝑛𝑐𝑒𝑠 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦 𝑖𝑑𝑒𝑛𝑡𝑖𝑓𝑖𝑒𝑑

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑖𝑛𝑠𝑡𝑎𝑛𝑐𝑒𝑠
Equation 3.2 Accuracy measure [45]
Evaluation of the performance of the Amharic braille recognizer is performed on two stages.
After the segmentation process we used manual count of dots to check the accuracy of
correctly segmented dots against the total number of dots in the braille sheet. In segmentation
we want to classify the image into dot area and non-dot area. But, due to error in the
segmentation two kinds of error could arise, a dot region recognized as a background and a
background recognized as a dot. Therefore, the performance measure considers the following
four results:-
A. A dot recognized to be a dot

B. A dot recognized to be background
11
C. A background recognized to be a dot
D. A background recognized to be a background
The point of interest in segmentation is the extraction of the dots from the background. For
such two class problem the use of a confusion matrix, as in table 1.1, simplifies the problem.
Predicted class
Dot None Dot
Dot A B
Actual
class None Dot C D
Table 1.1 confusion matrix
Therefore, from the confusion matrix we can set out that, accuracy is the sum of A and D
normalized against the sum of A, B, C, D.
We can use this same measure to evaluate the performance of the recognizer on character
level, which is after translation. To do this, we simply count the total number of characters in
the test braille sheet against the number of correct translations made.
1.5.5. Tools Used
MATLAB Image Processing Toolbox

MATLAB image processing toolbox is software that enables to perform image processing,
analysis, visualization and algorithm development. It has an extensive list of library packages
that allow performing image processing activities that an OCR tool applies such as image
property analysis and visualization, noise reduction, enhancement, color and contrast
adjustment, transformation and segmentation. One of its advantages is that it supports a
diverse set of images types. MATLAB also provides several products that can perform
together with the image processing toolbox software and extend its capability [36]. For this
reason, the interface building, preprocessing, segmentation and feature extraction processes
are done on MATLAB.
12
Python
Python is an interpreted object oriented and high-level programming language with dynamic
semantics. It has simple and easy to learn interface that emphasizes code readability. It is a
freely available, widely used general purpose programming language that supports modules
and packages [23]. Here in this work, python is used to map numbered feature class labels
with equivalent Amharic characters in a dictionary. Both operations, building of the
dictionary and matching against equivalent print characters, are performed on python. Python
is selected for the benefits mentioned above and for the ease of operation with Unicode
characters such as Amharic.
Hardware
The images used in this work as a dataset are scanned using HP flatbed scanner since it is
recommended by previous researchers. The Image processing and translation is done by
using the above tool running on ordinary windows computer with 2.6 GZ processer and 4GB
RAM. But, it is also possible to use computers with lesser capability but can satisfy the
lowest requirement for running MATLAB and Python applications.
1.6. Significance of the Study
Braille Character recognition in general allows digitizing hard coded braille documents in to
machine readable and editable texts. This can facilitate communication between visually
impaired and vision people. The other importance could be, it allows visually impaired
people to share their work to others and also enables preservation of historic braille
documents and can facilitate text-to-speech conversion.
In Ethiopia there are over half a million visually impaired people. Developing such a system
will benefit this people and any one working with them, such as family members, teachers
and public organizations. This research in particular will add to the search for Amharic braille
Recognition System, which to present time is not fully developed. There were few research
works on the area that searched for recognition of single-sided braille document. But, there
are a lot of braille document that are embossed in both sides. This research focused on
double-sided braille recognition for Amharic Documents and adds significant contribution to
the effort in the area.
13
1.7. Organization of the study
This thesis is organized into five chapters. The first chapter, Introduction, includes; the
background, the statement of the problem, objective, scope and limitation, Methodology and
Significance of the research.
The second chapter, Literature Review, presents review of previous works in the area of
Amharic braille writing and recognition, features of braille characters and characteristics of
braille recognition. It has also included general review of international researches in braille
recognition and double-sided braille recognition in particular.
The third chapter, Braille Recognition Systems, focused on the general process of braille
recognition with explanation techniques and algorithms used to develop the current system
involved in each step of recognition. This section will give emphasis to issues and main
techniques that can be used for separation of braille dots of double-sided braille.
The focus of the fourth chapter, implementation and experiment, is discussion activities
undertaken in implementation of the algorithms developed for the current system. It will also
discuss the results achieved and difficulties faced in implementing the current system. The
results are presented as recognition rate achieved both at dot recognition and character
translation level.
Finally, concluding remarks and recommendations for feature research are forwarded.
14
CHAPTER TWO
LITERATURE REVIEW
Braille has undergone through many modification and refinements in the past 190 years
before it gained worldwide recognition. Braille is a system with many advantages for the
users. It has great flexibility and ease of writing and recognition by touch. Any individual
can write Braille using a slate and stylus. These materials are relatively easy to get and as a
result nearly every blind person in every part of the world can enjoy the benefit of reading
and writing independently [42] [47].
2.1. History of Braille
Braille was not the first, or by any means the only method of touch reading. There were a lot
other means created before and even after the invention of Braille. The earnest desire of the
blind to find access to literature of sighted people opened the door for them which led to
many experiments in a variety of media [12].
The first ever recorded attempt to create script for the blind was made at about 1517, by
Francisco Lucas of Saragossa, who plotted a simple set of letters carved in thin tablets of
wood. This was later improved in Italy at about 1575 by Rampansetto of Rome, which
changed the nature of the writing into larger blocks of cut instead of raised. Nevertheless,
none of the systems were successful because of the difficulty in writing and reading them.
In 1651, Geoge Harsdorffer of Nuremberge revived the traditional method of writing on wax
tablet for use by the blind people through cutting the letters with a stylus. Twenty five years
later, Pardre Terzi devised a kind of cipher code based on a system of dots enclosed in square
and other figures and also an arrangement of knots tied in strings. There were also many
others who tried to create a method of writing and reading using movable raised letters of
lead, tin and cast metals. The problem with such systems was the letters were rough to touch
and hard to sculpt [12].
Another earlier effort includes the work of Maria Theresa Von Paradise, an Australian music
performer and composer, who assisted Valentine Hauy in establishing school for the blind,
devised a system of teaching the blind with the aid of pins stuck in cushions. But, when
Valentine Hauy founded his school in Paris in 1784, his first student, Francois Lesueur, found
15
by accident that he could feel the outlines of character “O” which had been strongly
impressed on a sheet of paper. Hauy later tried embossing books to teach his student. His
student made rapid progress and the system was a big success. However, it was the
innovation of this script by Louis Braille in 1825 which devised the system under which the
blind read to-day. But, it was not until some fifty years later that the Braille system was
universally adopted and, in the mean time, numerous other forms of embossed type were
planned and used; some employing lines and dots, others having the form of simplified
Roman capitals [12].
Braille got wider acceptance because it is more compact than any other system which
proceeded or followed it. It was outstandingly versatile, equally able to express the languages
and scripts of any nation. It has been readily adaptable to mathematics, musical notation and
other purposes. Its main advantages however lay in the fact that, unlike other embossed types,
it could be simply and easily written and read by the blind. It is a remarkably practical script,
preferred by the blind, which opened wide the gates to knowledge, Literary employment, the
ease to correspond privately with blind friends and wider opportunities for employment for
which the blind longed. Despite its manifest advantages, the general adoption of Braille was a
slow process [12].
Although, Louis Braille originally devised his writing for his expression of music, Braille
was adapted by many other countries that applied his system to various other practical
purposes such as the expression of mathematical and chemical symbols, the marking of the
faces of watches, meters, gauges, thermometers and playing cards and adapted too, to the
outlining of geographical maps and plans of cities and buildings [1] [12].
2.2. Braille Reading and Writing
Braille can be written in several ways. At the beginning it was written manually with stylus
which embosses the dots on strong thick paper framed into the slate to facilitate the printing
of dots. This raised dots formed in units called cells which are read with the finger tips of
blind people. But, currently, Braille is also written with Braille typewrites which can be
owned by many visually impaired and blind individuals for their day to day use. Recently,
technological developments in computing industry have provided additional means for
reading and writing Braille. It has offered software program and portable Braille note taking
16
devices that enable saving and editing of notes for use in places like lecture room and
meetings. These devices also enable displaying back either verbally or tactually. There are
also Braille embossers (Braille printer), which can be connected to any kind of computer with
the software and produce a hardcopy of a document in the same way ordinary printers work.
But, these devices are found only in some organizations with large requirement for Braille
production and duplication [1].
The most widely used Braille system consists of six dots and hence, sixty-three (26-1)
combination of dots can be produced which can be assigned for different characters. In the
six dot writing the dots are arranged in a matrix of three rows and two columns. The dots are
conventionally numbered. Those dots in the first column are numbered 1-2-3 from top to
bottom while the dots of the second column are numbered 4-5-6 in same manner. The
numbering is required to facilitate the description of individual symbols and for simplicity of
learning [12] [42]. For example, letter “A” is dot 1; letter “B” is dot 1&2; “C”, is dot 1&4 and
so on. The English alphabets, numerals and the most commonly used punctuation marks with
their Braille representation is shown is figure 2.1 below.
Louis Braille arranged the sixty-three Braille cell combinations into a symmetrical group of
seven. The first group of ten letters is composed solely from the upper four dots (1-2-4-5).
The second ten were formed by adding dot 3 to each of the first ten (1-2-4-5 + 3), the third
group of ten is again the same as the first ten, but this time with dots 3 and 6 added (1-2-4-5 +
3-6), while the fourth group of ten was once more the original line with dot 6 only added (1-
2-4-5 + 6). The fifth group is a repetition of the first line, but formed from the bottom four
dots (2-3-5-6) of the domino instead of the top four. The remaining thirteen signs were
composed of right-hand and bottom dots. The grouping of Braille cells is again believed to
facilitate Braille order for listing of Braille signs [42]. The complete order of the seven
Braille groups is shown in appendix I.
17
Figure 2.1 English Braille alphabets, numeral and punctuation marks
In English braille, the alphabets take up twenty-six of the sixty-three signs. The first twenty
letters are composed from the first two lines of groups. While the remaining five are formed
from the first 5 combinations of the third group and the letter “W” which was not found in
French alphabet is later represented by dots (2-4-5-6) from the fourth group. Ten are devoted
to international punctuation marks, while numbers are represented by the first ten letters but
preceded by a special numeral sign composed of dots (3-4-5-6). In case of writing large
numbers we only need a single numeral sign at the beginning. The normal Braille numbers
are used in sentences or page numbers only. For mathematical calculations, the modified
Braille notation, Nemeth code, which is not under the scope of this study, is used. The
18
remaining twenty-seven combinations are used variously to meet the special needs of
individual languages [12].
For a number of languages, two Grades of Braille have been established; Grade 1 and Grade
2 and in some countries cases Grade 3.
Grade 1 Braille
Grade 1 Braille, which is also called uncontracted, is the basic representation of Braille
symbols where a single cell represents only one letter, number, punctuation sign, or special
Braille composition sign. In this grade of Braille, we also have every letter of every word
spelt, letter for letter with the visual script. This grade of Braille is preferred for beginners
who are not familiar with contractions. There is also a language rule that disallowed the use
of contraction. The disadvantage of this grade of Braille could be that it is resource
consuming and as a result fewer books are transcribed in Grade 1[42].
Grade 2 Braille
Grade 2 is the everyday form of Braille writing used for general purpose Braille productions
of periodicals, books and letter-writing. It embraces a greater or lesser range of abridged
signs for the expression of conjunctions, prepositions, pronouns, prefixes, suffixes, frequently
recurring groups of letters and common words. In this system a cell can be used individually
or in combination with others to form a variety of contracted or short form of words. Its
primary purpose is to reduce the volume of Braille books which is in turn a saving in the cost
of production as well as in storage space and cost of distribution. It also saves the braillist
some of the effort involved in reading and writing [12] [42].
Few countries and languages have established highly abridged system, usually considered as
Grade 3 Braille, in which the original full text is scarcely recognizable and which borders to a
true shorthand. But, these are too complex for readers who are not endowed with three
qualifications; an extensive command of the language, an excellent memory and a highly
sensitive touch [12].
2.3. Double-Sided Braille
Braille has a very low information density. In average, a single page of 26 x 34 cm braille
sheet can only hold 27 lines and 32 characters per line [26]. This nature of braille has made
19
the production of documents in this writing to be voluminous and expensive. Many attempts
were made to overcome this problem. The introduction of braille contraction has significantly
helped this attempt of making braille documents concise. But, production of braille
documents was still bulky and expensive. Another effort taken was the introduction of
double-sided writing. Double-sided writing has made it possible to write on both sides of a
braille sheet the same way as the ordinary writing used by sighted people.
Double-sided writing or also known as interpoint is an offset printing or writing of braille to

allow embossing of the braille dots on both side of a braille sheet with the dots of one side
appearing between the dots of the other by shifting the dots on the other side. Double-sided
braille writing is found to be very economical to produce and a space saving option in braille
writing. It has also helped in reducing the size of bound Braille documents such as books.
However being economical and a space saving option, double-sided writing has created a
huge problem for OBR. This is because, in a given interpoint braille exists protrusions (dots
of the recto) and depressions (dots of the verso). And in a scanned interpoint braille image,
the dots of protrusions appear as a black region from above and a brighter region from below
while the depressions appear in reverse order. The major challenge in the recognition process
is that regions of the same type; for instance, white region form protrusion and white region
from depression, are merged creating confusion in the detection process [4] [50].
2.4. Amharic Writing and Braille System
Amharic has brought its own unique challenges for OBR attributed to the fact that it has its
own script that contains large set of characters. Therefore, understanding the unique nature of
the script is important in building Amharic Braille Recognition System.
2.4.1 Amharic Writing System

Ethiopia is an East African country with a population of more than 94 million people as per
World Bank 2013 statistics [17]. The country is home to many nations and nationalities
speaking more than eighty-four indigenous language. All the Languages in Ethiopia are
divided into four major groups known as Cushitic, Omotic, Nilo-Saharan and Semitic.
Amharic or Amarénya is the national language of Ethiopia as well a working language for
several regional states within the federal system of the country and as well for non-
government organizations and private institution. It is a southwest Semitic language of the
20
Afro-Asiatic language family. Being a Semitic language of the Afro-Asiatic root, it is related
with Hebrew, Arabic and Syrian languages [20] [33].
Amharic is believed to evolve from Geez which originated at around the first millennium
B.C. It is brought by immigrants from south Arabia who crossed the Red Sea into northern
Ethiopia, present day Eritrea, and mixed with local Cushitic languages and created Geez.
Currently Geez is not a spoken language but it is the liturgical language of the Ethiopian
Orthodox church. Amharic has also some affinities with Tigre, Tigrinya and the south Arabic
dialects. But Amharic has been highly influenced by local Cushitic languages, specially
“Oromo” and “Agew” languages, which can be seen from the non-Semitic words it used and
the sounds of the character sets of the language [20].
Amharic is spoken principally in the central highlands of Ethiopia. But, in terms of number of
speakers, currently it is one of the two major languages of Ethiopia along with “Oromifa”.
According to 2007 statistics 30% of the Ethiopian population speaks Amharic as a first
language and a roughly estimated additional 20% of the population use Amharic as a second
language. Therefore, it can be said that half of the Ethiopian population speak Amharic [17]
[33]. It is also known that outside Ethiopia Amharic has a large number of speakers in Eritrea
and it is also the language for millions of emigrants living in countries such as Egypt, Israel
and Sweden [20]. This fact had enabled Amharic to be the second most spoken Semitic
language in the world next to Arabic [33].
Amharic has a writing system called „fidel‟ which evolved from the writing system for Geez
which is known as Ethiopic or Geez otherwise. Ethiopic is a writing system which in turn
evolved from the old Sabaean writing system. The Ethiopic script has been in use since the
5th century B.C. But, the oldest extent records for Amharic are found only in songs and
poems dating back to the 14th century and literature in any significant quantity did not begin
until the 19th century [20].
Amharic Alphabet
According to Baye [21], there are three types of writing systems in the world, namely
logographic, syllabic and alphabetic. In the logographic writing system, a single symbol is
used to represent one word. It was first used around 5000 years ago near present day
Palestine and Syria. The old Chinese writing is also an example of logographic writing
system. In the second writing system, syllabic system, a symbol represents neither a word nor
21
a sound rather a phoneme, which is a combination of vowel and consonant. Therefore, the
number of symbols needed for a given language is determined by the number of basic sounds
used in the language. The third writing system, Alphabetic system, originated from the
syllabic system used in the Semitic languages of Middle East. It is also called the Greek
alphabet since its origin is strongly tied to the Greeks. The Latin alphabet is an example
which was originally known as the Greek alphabet before the Roman adapted it from Greek
and helped its wide usage in their Latin colonies [21].
Most African countries which were colonized by the west use the Latin script. Ethiopia is the
only African county that managed to avoid colonization by defeating the Italians during the
European scramble for Africa in the 19th century. As a result, it has not used the Latin
alphabet or other. It has its own writing system called Ethiopic. Ethiopian writing system
differs from the three writing systems discussed above. This system was originally a syllabic
system but, unlike syllabic, the vowels inside each phoneme are not represented by a number
of vowels but a single vowel [21].
The Ethiopic (Geez) writing system originated approximately 2500 years ago by the Semetic
Sabaean people of southern Arabia. The character sets are derived from Sabaean writing,
which has had twenty nine, according to baye[26], symbol in its unvocalized shape. It has
been in use for longer time in northern part of Ethiopia, particularly, Yeha until the Axumite
time whence it gave way to Geez. Amharic took the Geez script and become a written
language. It adapted all 26 Geez characters and modified seven of them by placing a bar/hut
on top of the characters and increased the number of characters to represent the additional
sounds acquired from Cushitic languages. Therefore, Amharic has now 33 core characters
where a single character appears in seven orders, one basic and six derived. Therefore, the
total number of syllables is 231(33*7) [21]. The non-basic forms are derived through
modifying the basic form by adding a circle or a dash at the top, bottom or middle position of
the basic character. There are also other symbols representing labialization, numerals and
punctuation marks. Figure 2.2 shows part of the Amharic alphabets. The full list of Amharic
alphabet can be found on Appendix II.
22
Amharic Characters
order Labialized
1st 2nd 3rd 4th 5th 6th 7th
ሀ ሁ ሂ ሃ ሄ ህ ሆ
ሇ ለ ሉ ሊ ላ ሌ ል ሎ
ሏ ሐ ሑ ሒ ሓ ሔ ሕ ሖ
ሗ መ ሙ ሚ ማ ሜ ም ሞ
ሟ ሠ ሡ ሢ ሣ ሤ ሥ ሦ
ሧ ረ ሩ ሪ ራ ሬ ር ሮ
ሯ ሰ ሱ ሲ ሳ ሴ ስ ሶ
ሷ ሸ ሹ ሺ ሻ ሼ ሽ ሾ
ሿ ቀ ቁ ቂ ቃ ቄ ቅ ቆ ቇ ቈ ቉ ቊ
በ ቡ ቢ ባ ቤ ብ ቦ ቧ
ቨ ቩ ቪ ቫ ቬ ቭ ቮ ቯ
ተ ቱ ቲ ታ ቴ ት ቶ ቷ
ቸ ቹ ቺ ቻ ቼ ች ቾ ቿ
ኀ ኁ ኂ ኃ ኄ ኅ ኆ ኇ ኈ ኉ ኊ ኋ
ኌ ኍ ኎ ኏ ነ ኑ ኒ ና
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . .
ጸ ጹ ጺ ጻ ጼ ጽ ጾ ጿ
ፀ ፁ ፂ ፃ ፄ ፅ ፆ
ፇ ፈ ፉ ፊ ፋ ፌ ፍ ፎ
ፏ ፐ ፑ ፒ ፓ ፔ ፕ ፖ
Amharic Numerals
፩ ፪ ፫ ፬ ፭ ፮ ፯ ፰ ፱ ፲
፳ ፴ ፵ ፶ ፷ ፸ ፹ ፺ ፻ ፼
Punctuation Marks
፡ ። ፤ ፣ ! ? ( ) " "
Figure 2.2 Part of the Amharic Script [53]
Amharic Numerals and punctuation marks
Amharic writing system has 20 signs for numerals. But, this numeral system cannot be used
for scientific calculations or complex mathematical operations. The problem is because the
Amharic numeral system doesn‟t have symbol for zero, decimal points, negative sign and
mathematical operators. Therefore, the Hindu-Arabic numerals and the Latin mathematical
operators are used for computational purposes. As can be seen in Figure 2.2 the unique
23
characteristics in the Amharic numerals are that they are enclosed in a top and bottom frame
in order to uniquely identify them from the alphabetic character sets [18].
According to omniglot.com [53], Amharic writing system has six punctuation marks; comma
(፣ ), full stop (፡ ፡ ), colon (፤ ), semicolon (), preface (:-) and question mark ( ) which is no
longer in use. But there are other punctuation marks which are adapted from Latin based
symbols like question mark (?), exclamation mark (!), quotes (“”) and parenthesis. The old
south Arabia monumental scripts regularly employ a vertical stroke as a word separator. This
style is adapted in Ethiopic texts as (፡ ) to separate words in a sentence.
2.4.2 Amharic Braille

Many countries have adapted and/or defined Braille code for their local languages. In Africa
Braille was introduced with the help of Missionaries and voluntary organizations during the
colonial periods. The oldest records of Braille appear for dozen African tribal languages such
as Swahili, Kikulu, kikmba, malgachi (Madagascar) Bemba, Chinyanja, Nyanja, Xosa,
Shona, Hausa, Mundang, Ibo, Twi and Kabili [12].
Amharic, along with Arabic and Afrikaans is one of the first three languages, which are
neither tribal nor linguistically African languages, used in Africa and have their braille forms.
Amharic and Arabic were the only languages having their braille notations built with their
own symbols. The rest of African braille was formed from European symbols such as
English, French, Norwegian and Dutch [12]. Amharic with a braille of comparatively recent
date is found in 1924 by American mission in western Ethiopia who together with other
voluntary and non government organizations established the first school for blind in
“Dembidolo” [12] [52].
The First version Amharic Braille was comprehensive and complete. It contained all Amharic
characters including characters with similar sounds, punctuations, numerals, mathematical
signs and musical symbols [52]. But, it was subject to many revisions. It was first modified in
1951 in consultation with Dr. A.N. Tucker, D.litt., Ph.D., M.A university of London, Miss
Hazel McGeery, American Presbyterian Mission, Sayo, and other workers for the blind in
Ethiopia and Sir Clutha Mackenzie, UNESCO Consultant on Braille [12].
According to the 1995 ENAB report [51], a revision was made once to eliminate redundant
Amharic Characters which have similar sounds while the other revision was made by other
missionaries to replace Ethiopic numerals with Arabic numerals and to further reduce the
24
variants of Amharic characters. Later in 2001 improvement was made considering issues set
by the World Blind Association (WBA) in adopting Braille to local languages. The current
version is fourth version which was published in 2001 by ENAB [51].
The Amharic Braille characters are composed of three parts, core and special letters,
numerals and punctuation marks [52]. As shown in figure 2.3 below in the fourth version,
each Amharic character has variants except the 6th order character. The 6th order character
does not require a vowel code to get the required sound of the character. Table 2.1 shows all
dots combinations for the vowel variants required in the 4 th version Amharic Braille.
Character 1st 2nd 3rd 4th 5th 6th 7th

Variant
Vowel 2:6 1:3:6 2:4 1 1:5 N/A 1:3:5
Table 2.1 Fourth version Amharic Braille Variants
Figure 2.3 Part of Amharic Alphabet Braille representation [46]
25
Amharic uses both Arabic and Ethiopic numerals. Combination of all dots (1:2:3:4:5:6) are
used to indicate Ethiopic numerals, while the combination of dots (3:4:5:6) are used to
indicate Arabic numerals. The fourth version Amharic braille numeral representation is
shown in figure 2.4 below.
Figure 2.4: Fourth Version Amharic Braille Numerals representation
Amharic Braille also consists of punctuation marks most of which are directly adopted from
English [32]. The only punctuation marks required are full stop and interrogation mark (2-3-
6). It is also to be noted that Braille dots should not be placed between the words as is the
custom visual Amharic text. Capital letters are not used; except for “ሯ” which will be
proceeded by Dot6. The 4th version Amharic braille representation for punctuation marks is
shown in figure 2.5 below.
Figure 2.5 Fourth version Amharic Braille Punctuation signs
This study will focus on the fourth version of Amharic Braille and sample Braille code for the
characters is shown in figure 2.3. The complete list of Amharic braille versions can be found
in Appendix III, IV, V and VI.
26
2.5. Automatic Braille Recognition
Braille recognition system is believed to simplify the communication between visually

impaired and vision society. This system can allow visually impaired people to access
literatures of vision society while at the same time enable them to share their works and to
preserve for future access. In recent years, a lot of researches are being carried out on the area
of Braille recognition locally and internationally.
Before converting a hard coded braille into a computer editable and printable text, an ABR
system goes through many processes. The most common operations are image acquisition,
preprocessing, braille dot detection, braille cell formation and interpretation. These main
processes are discussed here below in light of what researchers have done and recommended
on the area in recent years.
2.5.1 Image Acquisition

The first operation in any pattern recognition system is the acquisition of pattern data where
in the case of OBR would be the acquisition of braille image. For any optical recognition
process the quality of image captured has a huge impact on the recognition result. Therefore,
care should be taken while selecting the type of digitizing equipment to be used and on
parameter selection. The acquisition of images can be performed with the help of Scanner or
digital camera [13] [30]. Flat bed type of scanner is preferred by most researchers [3] [6] [22]
[30] [44] with an image resolution between 80 and 300 dpi [7] [11] [30] [37] [44]. The
resolution should not be lower so that important information about the dots will not be missed
and it should not be high so that unimportant details will not manifest.
2.5.2 Preprocessing
Preprocessing operations enable removing of noise and representing the scanned image with
only pixels which are relevant for the operation. It is mostly affects the classification
approach to be followed. Since different researchers follow different classification approach,
the type of preprocessing operation performed and the algorithms used by one researcher
vary from another. But, at this stage, such operations like conversion of full color image into
gray, detection and removal of noise, edge enhancement, morphological operation, and
skewed angle correction are commonly performed [31].
27
Computers store scanned full color image as 3D arrays. But, performing operations on such
images is very difficult and resource consuming. Therefore, the scanned braille image is
converted to gray scale image that represents the image in a 2D array [39]. On the gray scale
image [13] [44] applied a low pass Gaussian filter to reduce impulse noise and enhance the
edge, while [6] used an average filter followed by contrast enhancement to reduce the effect
of random noise.
Morphological operation enables enhancing of dots edges while skew angle correction
detects the angle in a wrongly scanned Braille image and corrects its degree. Skew correction
was not given attention by many researchers but there were some efforts by researchers in [3]
[1] [32] [35] recently. In [3] tried to perform a horizontal projection followed by number of
row count after rotating the image one time to left and one time to right. If the row count for
the left and right rotated image is the same the image is not skewed, else image is skewed and
iterative processing is done until the deviation between the left and right is more than 1/16.
2.5.3 Braille Dot Detection

Braille dot detection is the extraction of dot regions from the braille image. This is usually
undertaken by extracting the shadows created from the dots while scanning. In doing this, the
image is reduced to binary values. There are different methods to binarize a digital image.
One of these techniques, which is the simplest and most commonly used, is thresholding.
Thresholding applies a value that can classify the pixels into different regions of intensity.
Thresholding can be global or local. Further, the threshold value can be single or multiple. In
using thresholding for image binarizaton, the main problem is finding an optimal threshold
value [19] [26] [31].
Two types of thresholding, global and local, can be applied in braille image segmentation [8]
[31]. Global thresholding calculates a threshold value from the total intensity level of the
pixels that make up the image. Then every pixel in the image will be compared against this
threshold value and its class value will be set. Local thresholding calculates threshold value
after classifying the image into several areas. This will result in different threshold value for
different regions of the image. Then every pixel will be compared against the locally
computed threshold and its value will be decided [19] [26]. Local thresholding performs
better in cleaning out noise and separating the area of dot from background [31] [34]. In [37]
an innovative thresholding technique that uses an iterative algorithm that looks for pixel
28
value from an optimum area of Braille spots is applied. In [7] used a statistical method based
on beta distribution of a gray scale image to find an optimum threshold values.
2.5.4 Braille Cell Formation

On the threshold image, algorithms are applied to extract areas of dots and group the dots into
cell. The most commonly used algorithms used mesh grid formation and local measure. In [7]
used grid formation by selecting a starting point and then generating a horizontal and vertical
line on a regular interval. Separate grids are formed for the recto and verso dots by checking
the color distribution. This technique can help in removing of wrong dots and recovery of
possible dots. But, it has problem likes difficulty of identifying dots which can be connected.
In addition the whole performance could fell if the scanned Braille document is not correctly
angled [3] [7] [32]. [3] Performed a horizontal pixel search for part of dot using 8 pixels
height for dot. The 8 pixels are composed of a dark and a bright region with a space between
them. Pixel values in the image with such property will be identified as dot parts and based
on the color distribution a global measure will identify the image into columns of recto, verso
or undefined column. But, [16] used a little different approach, that use a predefined cell
frame size of 95X80 that can move over the image to check for Braille cell and when found
the dot size of 20X15 pixels will be used to check existence of dot on the possible regions
inside the frame. However, this technique may not be good for double sided image.
Therefore, [25] applied a grid inside the frame that classify the frame into three and check
existence of recto or verso dots using the Braille distance rule.
2.5.5 Interpretation
After all dots are detected and grouped into cell, a binary code is created for every cell based
on the location of dots in the cell [25]. In the binary code a value of 1 is attached where dot
exists and 0 otherwise. These binary codes can be converted into decimal values and
interpretation to ordinary text can be performed by comparing the values with a stored
dictionary or lookup table [3].
2.6. Review of Braille Recognition Works
2.6.1 Review of Amharic Braille Recognition works

The attempt for applying OCR technology for Amharic Braille documents was started
recently. The recent research works on the area that attempt to find good techniques for
29
recognition of Amharic Braille Characters includes the works of Teshome [43], Ebrahim
[18], Shumet [41] Miftah[32] and Berihu[9].
To the best of the researchers knowledge Teshomes‟ [43] work was the first attempt in
developing Amharic Braille Recognizer. To acquire the image of the Braille document he has
used a regular flatbed scanner with 200 dpi. Braille images were binarized using global
tresholding technique which has shown better computational performance than local
tresholding. To segment the dots from the image, Teshome [43] used mesh-grid construction
and adopted a modified region based approach to extract features. The features extracted are
used to build a model, using the MATLAB feed forward Artificial Neural Network (ANN),
which is used for Braille character recognition. The experiment has shown that the designed
system can perform with 92.5% accuracy.
Teshome [43] performed his experiment on clean Braille document. However, in real life
application of Amharic Braille recognition, it is essential that the system recognize a noisy
Braille. This is because Braille document by their very nature are damaged due to repeated
use, bend and scratch. This will create noise on the image of the Braille. This noises and
impurities highly affect the recognition result. Ibrahim [18] aimed to explore different noise
detection and removal algorithms in order to enhance the performance of the Amharic Braille
recognizer in real life Braille document images. With this regard, Ibrahim [18] examined and
identified attributes and features of Amharic Braille characters and selected the best noise
detection and removal algorithms. After this, a program is developed for the noise removal
algorithm and is integrated into the Amharic Braille recognizer.
The system was implemented using ANN for 238 Amharic characters, 10 numerals and 19
punctuation marks. Ibrahim [18] used Gaussian filtering with adaptive histogram equalization
and morphological operation to detect and remove noise in real life Amharic Braille
documents. The noise detection and removal and thresholding algorithms are integrated with
the previous algorithms developed, by Teshome [43], for clean Amharic document and the
system is able to achieve an accuracy of 95.5%, 95.5%, 90.5% and 65% for clean, small level
noisy, medium level noisy and high level noisy documents respectively.
The effort made by Shumete [41] was to enhance the Amharic Braille Recognition system by
focusing on feature extraction and classification. With this regard, three feature extraction
algorithms: fixed cell measure, horizontal and vertical projection and grid construction were
experimented. The test result shows that the fixed cell measure performs very well. To build
30
a classification model for prediction of Amharic characters from Braille cell, J48 decision
tree and the support vector machine (SVM) classifiers were tested. Based on the experimental
results, SVM outperforms decision tree classifier in predicting unseen extracted Braille
features. After application of the explored feature extraction and classification techniques to
real life Braille documents, an overall accuracy of 90.67% was achieved.
Miftah [32] presented an approach for recognition of Amharic Braille document which
focused on three major areas; noise detection and removal, Braille dot recognition and skew
correction. The main focus of his work was the application of direction field tensor, that uses
Gaussian filters and derivative of Gaussians for noise removal and isolation of Braille dots
from their background. The other contribution of this work was skew correction which
exploits the horizontal direction nature of Braille dots. Other contribution of this work
includes the construction of Braille character lines to allow subsequent operations to be
performed only on the character line. A half character detection method which helps
differentiate dots from noises is applied. After detection of half characters, a horizontal
distance analysis between the half characters is performed to find complete Braille cells.
After formulating the Braille cells, a lookup table is used to translate them into corresponding
Amharic characters. This system has achieved 99.9% and 96.5% accuracy for good quality
and poor quality Braille documents respectively
As a continuation to previous works in the area of Amharic OBR the main focus of Berihu
[9] research was to add Tigrigna Braille characters to the already designed Amharic OBR and
to design a generic Amharic - Tigrigna Braille recognizer that can handle the various artifacts
like dot size, merged dots and alignment of dots in Braille documents that includes designing
of a noise tolerant segmentation technique and translation algorithm that helped to improve
performance of the system in recognizing Amharic-Tigrigna real-life braille documents. The
system was tested with ANN and SVM classifiers but it has achieved better result in using
lookup table. The average accuracy level registered is 96.5%.
As we observe from the above research works, all are focused on single sided Amharic
Braille recognition. But there are research works that attempted to recognize double sided
Braille for English, Chinese, Arabic and Spanish.
31
2.6.2 Review of double-sided Braille recognition works
Regular feature extraction for recognition of Braille
In [13] presented an automatic system for recognition of both single-sided and double-sided
Braille pages and converting to an editable English or Chinese text document. Before the
system gets the Braille document from the scanner, consideration was made to the
environmental factor in order to minimize the problem of image analysis. With this regard,
the embossed Braille page is illuminated with a yellow polarized light source placed at an
angle of 45 degree away from the page top and the dots of both sides are able to be captured
clearly in the scanned image. The images are captured with a digital camera at a resolution of
512*512 pixels. On the captured image, two pre-processing operations are performed; noise
filtering and edge enhancement. To filter the noise, a low-pass spatial Gaussian filter is
applied that clean out the high spatial frequency introduced in capturing the image. But, in
addition to filtering noise, the low pass spatial Gaussian filter has helped preservation of the
detailed edge information of the Braille dots. The edge enhancement operation is performed
to sharpen the fine details of the image that has been blurred. With this regard, two
independent filtering operations using convolution Sobel kernels are applied.
For segmentation, the researchers considered a histogram made up of only those pixels that
lie at or near the edges of the Braille dots. The pixels that lie on or near an edge are
determined by using the Laplacian of Gaussian operator. The average gray-level value of
these edge pixels are then used as the global threshold value. This has improved the
segmentation result. Before feature extraction the boundary points of Braille dots are
obtained using boundary based chain code algorithm which detect the boundary coordinates
of the Braille dots. In case of double sided writing, the dots on both sides should be separated
before any further processing. Using the boundary coordinates and illumination
characteristics, two standard templates that represent front-faced and back-faced dots are
constructed. Then the templates are applied to every pixel position of the image and the
correlation at each pixel position is evaluated. Based on the correlation value obtained, the
separation of the Braille dots is performed. In order to group the dots into cells, the distance
between the centroid and the four possible neighbors are determined. Then within the cells,
the dots positions is determined and represented as a bit string. Finally, the bit strings of the
cells are searched against a Braille dictionary and retrieved characters are grouped into
words.
32
This work had shown that recognition process starts before image acquisition and application
of a yellow polarized light source from an angle of 45 degree has trivialized the image
processing work. The edge enhancement technique applied has improved the image very
well and the segmentation technique applied was highly effective. In addition, the
application of a post recognition activity has helped to improve the recognition output. As a
result, the system has achieved a recognition accuracy of 100% for single-sided Braille and
97% for the double sided Braille.
An Application of Eight Connectivity based Two-pass Connected-Component

Labeling Algorithm for Double Sided Braille Dot Recognition
The objective of the system proposed in [48] is to develop an optical Braille character
recognition system that takes into consideration differences in scanned image quality and
resolution. First 24bit images of Braille are captured with Hp Scan jet 3400C A4 size scanner
with an image resolution of 200dpi and spatial resolution of 1501 x 2121. The algorithm
begins by converting the captured colour image to gray scale image. In the scanned image, it
is observed that the shadow produced by the depressions relatively account for a lesser
number of pixel count when compared to those produced by the protrusions. This is due to
the reflection property of the light. Therefore, it is believed that retaining only the brighter
areas can allow recognition of both sides of dots at once.
In order to remove any inherent noise present in the image, thresholding is performed on the
gray scale image. A single thresholding is applied where pixel values less than the set
threshold are replaced by zero and values greater than the threshold are retained as they are.
But, in the threshold image an impulse noise is observed. To remove the impulse noise a
median filtering is applied. Further morphological dilation is performed in order to increase
the area of the dot. The morphological dilation is more helpful in increasing the area of the
verso dot rather than recto dot.
The main purpose of this project is separation of verso and recto dot. With this regard, the
eight connectivity property of the pixel relationship is employed. Eight-connectivity for any
pixel means that it is connected horizontally, vertically and diagonally. When applying the
eight connectivity pixel operation, the count of the recto is found to be less compared to the
verso. Therefore, differentiation between the recto and verso dot is done by applying
thresholding operation on the basis of eight connectivity based pixel count value. Those dots
33
with eight connectivity count greater than the threshold are considered as the verso dots and
those dots for which the eight connectivity count value is less than the threshold are
considered as recto dots. The separated dots are then placed into separate document in a
location similar with the original document. After separating each side, the Braille dots are
transcribed into their corresponding natural text. In order to do the transcription it is required
to group the dots into cells first. Usually grouping dots is made by constructing grid over the
dots with support of the standard Braille dimension. This technique is ineffective and time
consuming. Therefore, this paper proposed the use of adaptive grid construction technique
which is found to reduce computation time and is also effective in avoiding any
misclassification of dots. This is possible because the adaptive grid construction technique
applies an algorithm that considers three factors of Braille image; distance calculation,
horizontal projection profile and vertical projection profile.
During the dot recognition process all the valid Braille dots have been detected on either
sides of the document and the two images are formed for each side of the document. To
convert the recognized dots into their corresponding natural characters, the scanning pattern
is used for dividing each cell into grids consisting of six parts and corresponding code for
each cell is generated according to the presence or absence of a dot in each grid. The presence
or absence of all the valid Braille dots is found by multiplying the grid constructed image
with the dot extracted image. If the product is true then it indicates the presence of dot and its
value is indicated with 1 and if the product is false then it indicates the absence of dot and its
value will be 0. The dot positions are determined with number 1 to 6 as per the universal
numbering of Braille dots. Within each cell, the dot pattern is determined and is also
represented by a bit string which is later converted into equivalent decimal codes. To retrieve
the natural characters corresponding to the Braille characters, a matching algorithm is applied
in which, decimal code generated from the processed image are searched against
corresponding Unicode characters stored in the lookup table. Then the MATLAB function,
unicode2native, is used to convert the Unicode character into corresponding natural text.
This paper presented a novel technique that make use of an eight Connectivity based two-
pass Connected-Component Labeling algorithm and thresholding technique with an adaptive
grid construction to separate and recognize double sided Braille. The proposed system was
able to separate and as well recognize double sided Braille document with a minor error. The
errors happened due to failure to handle all noises in Braille image. The proposed system also
failed to handle deformation, skew and merging of verso and recto dots.
34
Braille character Recognition using Generalized Feature Vector Approach
The main objective of [25] is the application of a multi-parameter based feature detection
algorithm along with a sliding window and a grid to efficiently detect the embossed dots from
a Braille print. Based on two steps model selection process, model is selected from the
segmented image and parameters are calculated. Second, on the image segmented with colour
parameter further calculation is made for Generalized Feature Vector.
On the acquired image a portion is select and on the segmented image a model is selected and
parameters are calculated. To detect the dots, a sliding window with a fixed interval is used.
The size of the sliding window is decided considering dot dimension measure of print Braille
and the standard inter-dot and inter-cell spacing. The dot position for each cell is then
detected using GFV. GFV is a multimodal probability distribution in a multidimensional
feature space which essentially encapsulates multiple features for consistent detection. It
enables the use of multiple features to reduce error and false alarms or wrong identification.
Each cell is divided into grids and the presence and absence of a dot is sensed and
corresponding code for every cell will be generated. In the case of double sided Braille, the
algorithm is further modified to enable it distinguishes the front side dots alone from the
Braille image. The process is similar, using a sliding window, but uses a grid method to
distinguish between the front and back side. The sliding window will be sub-divided into
three vertical regions. The front sided dots will be extracted based on their location in the
grids. For example, a dot fully appearing in region 1 and 2 is a front side dot. A dot appearing
partially in any of the regions is backside dot or a false dot created when dots from the two
sides are merged. In addition to the sliding window and the grid, here the GFV parameter
which consist size ratio analysis along with color and energy matching will be applied to
differentiate between a single dot and a merged dot. Finally, after dots are extracted, a code
based on the position of dots in a cell will be generated. The generated binary code will be
matching up with a lookup table.
The unique operations of the proposed system are the selection of sliding window dimension
and interval which will enable extraction of cell position before dots that make up the cell.
The other novel technique is the use of generalized feature vector for detection of dots. The
researchers claim that the GFV approach can efficiently detect Braille dots even from low
quality image. Though, the accuracy of the proposed system is not quantified, it is shown that
35
the system can successfully distinguish between front and back side dots. One of the
weaknesses of this system is that, simultaneous recognition is not possible.
The present work made an effort to recognize both single-sided and double-sided Amharic
braille documents. In the current work, an effort is made to improve the image acquisition
process by considering the environmental factors that affect the scan quality in order to
minimize the problem of image analysis. This work has introduced a combined technique that
uses both projection profile and grid construction to segment and extract features of braille
dots using only those dot regions which are perfectly constructed. It also uses a grid
improving technique that is found to reduce errors to minimum. As a result, it can be said that
the current work highly reduces the process of image analysis but with improved result.
36
CHAPTER THREE
AMHARIC BRAILLE RECOGNITION
Vision is the one sense that we come to depend upon above all others and indeed the one that
provides most of the data that we receive from our environment. The eye which is the input
pathway of visual information from the outside world into the brain provides two times
greater information to the brain than all other senses collectively. Though the process
between the eye and the brain is very complex, the visual perception and interpretation
activity is normally carried out in tenth of a second. In the past few decades people have been
trying to get machines do much of their works. Machines have been doing simple
mechanistic tasks with no difficulty. But, for more complex tasks such as those that require
sense of vision, the thing that is trivial for humans has been very challenging for machine to
perform and computationally demanding. This fact is evident in character recognition
systems [5] [19].
Character recognition systems are used to convert characters in hard coded documents into
computer readable and editable text. One of such system is Braille Character Recognition
which is a system that reads braille documents and converts the braille characters into their
equivalent textual representations. This recognition process is dependent on image
processing, artificial intelligence and computer vision capacity [29].
Braille recognition can be either single-sided or double-sided based on the way the braille
document is prepared. Unlike human recognition that recognizes objects as they are, optical
recognition systems work by deducting every object separately from the background [19].
This task makes recognition in general and that of braille very difficult. Specially, with
double-sided documents where the dots of both sides are visible from one side, the dots will
be highly connected to each other and separation of a single dot is a challenging task [26].
The whole braille recognition task is divided into five main operations; image acquisition,
preprocessing, segmentation, feature extraction and translation. The recognition process starts
by acquiring image of the braille document through a scanner, digital camera or any other
digitization means. The digitized images are preprocessed to filter out noise and unnecessary
information which are not relevant for the recognition process. Various techniques are
applied to segment braille dots from the preprocessed image and form a braille cell. Finally,
37
features of the segmented objects are used to translate them into the desired language textual
representation.
This research is concerned with the application of different braille recognition techniques for
conversion of double-sided Amharic braille documents into Amharic print characters. For this
purpose, various approaches used in braille recognition such as image digitization,
thresholding, segmentation, feature extraction and recognition are reviewed.
3.1. General Design of Amharic Braille Recognition
The overall design of the Amharic Braille Recognition system is shown in figure 3.1. The
main focuses of this work are highlighted with double rectangles.
The system is able to recognize both single and double sided Amharic braille documents. To
translate Braille code characters into their equivalent textual representation a series of steps
are required. In every character recognition system including braille recognition systems, data
is feed to the system in a form of an image. Therefore, sample Amharic braille documents are
collected and an image of the braille document is acquired using a flatbed scanner with 200
dpi. The acquired braille color images are then converted to gray scale. On the gray scale
image preprocessing operations such as noise filtering, intensity adjustment, size reduction,
thresholding and morphological adjustments are performed. Before binarizing the image for
segmentation, separation of the recto and verso dots is performed using a moving window
filter. The front and back side dots are separated and put into two binary images. The
binarized braille images are feed to an improved segmentation and feature extraction
techniques. Finally, a lookup table is developed based on which the extracted features are
parsed to their equivalent Amharic textual character representation.
38
Figure 3.1 Design of Amharic Braille Recognition System
39
3.2. Image Acquisition/Digitization
Any character recognition system starts by acquiring the document in a form of an image.
The acquisition process can be done with the means of digital camera or scanner. A flat bed
type of scanner is the most preferred technology by many researches. This is because flat bed
scanners are cheap alternatives that can be operated easily and produce the required image
with minimal time. In OBR also, scanners have been in use dominantly.
To be able to distinguish the dots from the background, the recommended spatial resolution
setting is between 80 and 200 dots per inch (dpi) [7],[16], [26],[30]. Other things to consider
while scanning a braille document is that the color of the document to be scanned can affect
the output. This happens due to the effect of the reflected light from the scanner during the
process. We have observed that such effects can be handled by setting scanner to apply
adaptive light mode, rather than the default settings that apply constant lighting for any
colored document type. But, this option may not be available on every scanner. Another
option can be scanning in gray scale or reducing contrast and brightness during scanning. In
this work an adaptive lighting is used to get a quality color image of the braille documents.
3.3. Preprocessing of double-sided Braille Images
In OCR, image preprocessing is an important step which includes the activities of converting
RGB image into gray, filtering image to reduce noise, skew angle correction and many
others. After pre-processing the image, only important regions of the image will be extracted.
This process is called Thresholding [19] [38]. Thresholding, in braille recognition, enables
the extraction of Braille dot regions from the background from a filtered gray scale braille
image. In double-sided braille recognition, only three regions can be useful; light area,
shadow (dark) area and background area. Therefore, it is enough to represent the images into
three color levels, White, black and average gray, rather than full 255 gray levels. The dot
areas appear as regions with combination of black and white, while the background will hold
an average gray value of the image.
There are different threshold values setting techniques applied by different scholars for braille
recognition. In this work, the thresholding technique used by [3] is adopted with some
modification. The thresholding technique is detailed in algorithm 3.1 below.
40
1. Calculate the average gray level for the whole image
Let image be an array representing the whole image
Let x and y be the horizontal and vertical coordinates
set x=y=0; // set coordinates to first pixels
for x=0 to n rows in image do
for y=0 to m columns in image
sum+=image(x,y)
while end of image
avg=sum/(n*m)
2. calculate values for a and b
Let max(image) be highest value in each column
Let min(image) be lowest value in each column
a=mean (max(image)+ avg)/2;
b=mean (min(image)+ avg)/2
3. add values higher than a to array y and values less than b to array x.
4. Calculate the max_avg and the min_avg
max_avg=mean(y);
min_avg=mean(x)
5. Compute the classification parameters, h and l
h = avg + (max_avg – avg)/3;
l = avg - (avg – min_avg)/3
6. Set Pixel values higher than h to 255 (white)
pixel>h=255
7. Set pixel values less than l to 0 (black)
pixel<l=0 and
8. Set pixel values between h and l to mean(average)
h>avg>l
Algorithm 3.1 Thresholding algorithm [3]

This thresholding technique calculates two values, higher and lower bound, from the image
that classifies the image into three regions. Any pixel that lie below the lower threshold value
is set to 0 (black), any pixel that lay above the higher threshold value is 255 (white) and any
pixel in between the two threshold values is considered as background and is set to average
pixel value of the image.
3.4. Segmentation
The thresholding operation leaves us with an image having three intensity levels only; black,
white and an average gray color. The black and white regions represent the foreground (dot
regions) while the average gray region is the background. This image can hold both recto and
41
verso dots. Therefore, before any other operation is performed, the regions of recto and verso
dots should be identified. After the regions of the dots are extracted, we separate the recto and
verso regions into different array. For each image, image holding recto and image holding
verso, we can extract the dots from the region identified and later group them to form a cell.
3.4.1. Braille Dot Detection

In the preprocessed image, a single dot appear as two separate regions and in the case of
double-sided writing, the dots from both sides are highly connected. This makes detection of
a complete dot at once difficult. As discussed in [3], detection of dot parts rather than
detection of whole part produces better result. As per observation from the scanned image,
the size of the dots from the back side is smaller as compared to those from the front.
However, from experiment, the average dot height is found to be 8 pixels for recto and 7
pixels for verso. This information is used in the algorithm that separates the dot region from
the background.
3.4.2. Separation of recto and verso

The recto and verso dot separation is performed with the help of the black and white region
orientation. The black and white regions are formed as a result of the shadow created from
the scanner in the process of acquiring the braille image. As discussed above, in the threshold
image, the recto dots appear as white region from above and black region from below, while
the verso dots appear as black from above and white from below. Sample, recto and verso
dots can be seen in figure 3.1 below. In a double sided braille image, the dots are highly
connected. The connection, most of the time, happens when similar region from recto and
verso are connected. This way two or more dots can be connected together which makes the
separation of a single dot from the image very cumbersome.
(a) (b) (c)

Figure 3.1 Images of recto and verso dots a) recto dot, b) Verso dot and c) Recto Verso dots
connected
Initially, an effort is made to separate connected dots using morphological operations. But, in
the resulting image, the shape of the dot was highly affected and it was not successful in
separating all connected dots. As a result recognition performance was very low. In this
42
work, we have used a moving window filtering technique which is adopted from [3] and
modified to enable recognition of frontal side and backside dots without the need to separate
connected dots. The technique is explained in algorithm 3.2 below. The algorithm is designed
in such a way that it does not scan the same area twice and it skips 8 pixels when a recto or
verso region is found. This has reduced computational time and number of false recognitions.
At the end, we have two arrays holding the dot regions for the recto and verso dots
separately.
1. Create a matrix w of size 8x1 with the first four of its value 255 and the last four 0
2. Create two empty arrays that will hold the recto and verso dots separately
3. Start from the first pixel and perform a vertical scan with a window size as w
4. Compare the selected window against criteria
Criteria:-
a. first two or three pixels are bright(255) while last two are dark(0)
b. first two pixels are dark(0) while last three are bright(255)
5. While criteria ‘a’ is true
- multiply the selected window with w to get the bright region only
- register the result to recto array
- skip 8 pixels vertically
-go back to step 4 until end of image
6. While criteria ‘b’ is true
- multiply the selected window with w but this time w fliped vertically to get the
bright region only
- register the result to verso array
- skip 8 pixels vertically
-go back to step 4 until end of image
Note:- the value 8 which is used as a window size to filter the image is selected with
experiment to be the average dot height.
Algorithm 3.2 Front and back side dots separation algorithm [3]
A single dot, on average, appears with 3 pixels height and 3 pixels width. But sometimes, due
to the quality of the scanned braille image, a dot region may not be detected at all or may be
detected with more than one separate region. In order to recover such improperly detected
regions, a neighborhood analysis is performed that will look if a pixel‟s horizontal neighbors
43
are a dot region. If the case is true, that pixel will also be considered as a dot region. This
technique can also help to remove most of the wrongly detected regions; since those regions
mostly appear as an isolated one column region, the neighborhood pixels constitute a none
dot region. This and other morphological operations had improved the shape and size of the
dots detected.
3.4.3. Braille Cell Grouping

After the dots are detected, they are grouped into cell that makes a single character or part of
a character. Having identified all possible dot regions and separated the image into recto and
verso, we have two separate binary images to be processed; image holding the recto dots only
and image holding the verso dots only. However, the verso image is flipped horizontally to
bring it to its normal reading position. As discussed in [41], there are different techniques for
grouping braille dots into cell such as fixed cell measure, horizontal and vertical projection
and grid construction. All this techniques try to make use of the fixed matrix arrangement
nature of braille dots.
In this work, an improved flexible grid construction technique that uses horizontal and
vertical projection results to find the coordinate value for the grid is applied. First, for each
separate image, we find the horizontal and vertical projection profile. But, building a
horizontal and vertical projection on all dots in the image will produce incorrect result. This
mostly happens due to two reasons:
1. Over projection:-
When dots which are smaller in size are separated in distance from other dots in the
line, they create their own projection. This will create horizontal or vertical projection
beyond the number expected.
2. Under projection:-
When dots are bigger in size they will fill up the vertical and/or the horizontal space
that should exist between dots. In this case, the horizontal or vertical projection will
be less than the number expected.
Therefore, to alleviate these problems, a dynamic algorithm is applied that extracts properly
sized dots only and build the horizontal and vertical projection for the whole image from
those selected dots only. The grid is constructed from the horizontal and vertical projection
profile. The coordinate values to construct the grids are calculated from the centeroid values
44
of each connected component in the horizontal and vertical projection array. As discussed
above, the size of braille dots is not uniform. Horizontal and vertical projection works on an
area that encloses every dots region such that no dot exists outside this region. This is
computationally ineffective and resource consuming. But, using the centeroid values from the
projection profile enables us to get a line that crosses most of the dots in that region. The
technique is explained in detail in the algorithm 3.3 below.
1. Calculate the area of every object in the image using

Let n be total number of objects in image
For i=0 to n do
Calculate area
2. Calculate the average area of the dots
-Let m be matrix holding area of every object
Average_area=average(m)
3. filter object with area between average_area - 2 and average_area +2
let s be objects filtered
(average_area -2)>s> (average_area +2)
4. Create a vertical projection array which is the size of the new image height
and the sum of connected pixels in every row
5. Create a horizontal projection array which is the size of the new image width
and the sum of pixels in every column
6. Compute centroids of every connected component in the projection
7. Connect centroid point from vertical projection to connect pixel in y axis
8. Connect centroid point of horizontal projection to connect pixel in x axis
Algorithm 3.3 Grid construction algorithm

When building grid on a perfectly extracted braille dot region with the centroid values of the
horizontal and vertical projection results, the line will perfectly fall over the center point of
the dots. The vertical lines in the grid appear as two parallel lines while the horizontal line
appears as three parallel lines.
On a well constructed grid image, any possible cell region can be reached. This is possible
because, when three horizontal parallel lines are crossed by two vertical parallel lines they
create six intersection points. The braille cells are extracted by simply looking for any six
intersections with smallest distance between them. The six intersection points will be the
45
position of the six dots in a braille cell. But the above algorithm finds grid only where there is
at least one perfectly sized dot in the region of the image. In order to have a complete picture
of the cell, it is needed to have grids on possible dot areas as well. This problem is solved
using an algorithm that refines the grid by checking if a gridline exists in all potential dot
areas and draw a grid where not exists. From the constructed grid, the algorithm finds two
and three consecutive and perfectly constructed vertical and horizontal line that represents the
two columns and three rows in a cell respectively. The space between these lines is used to
estimate the other lines that should exist in the image. This technique is detailed in algorithm
3.4.
1. Refine horizontal grid line

a. Loop through all line and find three consecutive lines having similar
space between them.
b. Calculate GH using estimated vertical space between dots (vsb) and
estimated vertical space between cells(vsd)
GH=vsd + vsc
c. Take the coordinate value of those selected lines
d. Starting from the selected lines move iteratively through the image, both
towards top and down, in step of GH to check if a grid exists on or near
the point. The checkpoints are the potential dot areas.
e. Add new horizontal gridline where none is identified on the check point.
2. Refine vertical grid
a. Loop through all line and find two consecutive lines having average
horizontal dot space between them.
b. Calculate GV using estimated horizontal space between dots (hsb) and
estimated horizontal space between cells (hsc)
GV=hsd + hsc-
c. Take the coordinate value of those selected lines
d. Starting from the selected lines move iteratively through the image, both
towards left and right, in step of GV to check if a grid exists on or near
the check point. The checkpoints are the potential dot areas.
e. Add new vertical gridline where none is identified on the check point.
Algorithm 3.4 Grid refining algorithm
46
3.5. Feature extraction
In this work, to extract features of a given cell, the existence of dots on or near the
intersection points of the grid is checked. First, an array is built, to hold the feature of the
image. By starting from the left top corner, the array is populated with either zero or one
based on the existence of dot on the intersection points. One implies there is a dot on the
intersection point, while zero implies there isn‟t. The scan is made in such a way that it will
scan one cell first in the order from the first dot to the sixth and move on to the next until the
end of the grid. To do this, the algorithm start from the first intersection point in the grid,
move on horizontally to the next intersection point in the same vertical line and next to the
third intersection which is again on the next horizontal line on the same vertical point. This
makes half a cell. To complete the other half, we move to the next vertical point on the same
horizontal line as the first intersection and do the scan the same way we did for the first half.
The method for extracting the features is detailed in algorithm 3.5 below.
1. Create an empty array with the size of the intersection points in the grid
2. Take the first three lines in the horizontal grid which represent the three
lines in a single cell,
a. Start on the first intersection point and check if there exists dot
b. Move to second line of the first three and check the same
c. Move to the third line of the first three and check the same
d. Move back to the first line and perform a check on the second
intersection
e. Jump to the second line and perform a check on the second
intersection
f. jump to the third line and perform a check on the second
intersection
This make up six checkpoints which represent the six dot positions in a cell.
3. For every checkpoint, if a check results true put 1 on the array else put o.
4. Scan the remaining intersection points on the selected horizontal lines of
the grid the same way
5. Move to the next three lines in the grid perform step 2-4 until end of grid
Algorithm 3.5 feature extraction algorithm
47
At the end of this process, we have the feature of the cell in an array of zeros and ones. The
size of the array is six times the possible cell positions identified by the grid. Braille cells are
represented with six binary values that represent the existence of dot on six possible dot
positions. To make up a binary code value for each cell, we look for any six consecutive
binary values starting from the first point and moving in a step of six in the array holding the
features. After we get the binary value of each cell, it is converted in to decimal code
representation for simplicity of subsequent operation which is translation [3]. As shown in
formula 3.1, the decimal code is the sum of the multiple of the six binary values with six
decimal points; 1, 2, 4, 8, 16 and 32 in the order from the first dot to the sixth. The six
decimal points imply the coefficients for converting a binary (base 2) number in to decimal
(base 10) number.
𝑵
𝒏𝒖𝒎𝒃𝒆𝒓-𝟐 = 𝒅𝑵 … 𝒅𝟐 𝒅𝟏 𝒅𝟎 𝒏 = 𝒏=𝟎 𝒅𝒏 𝒃𝒏 = 𝒅𝟎 𝒃𝟎 + 𝒅𝟏 𝒃𝟏 + 𝒅𝟐 𝒃𝟐 + ⋯ + 𝒅𝑵 𝒃𝑵
Where b - Numeral System base
dn - the nth digit
n - can start from negative number is number contains fraction
N+1- Number if digits
Equation 3.1 binary to decimal conversion rule
For example a braille cell that have dot on the 1st, 2nd and 5th position will have a binary
code of 1,1,0,0,1,0 and its decimal value can be calculated as:
Decimal code (base 10) = 1*20 + 1*21 + 0*22 + 0*23 + 1*24 + 0*25
Decimal code= 1*1 + 1*2+ 0*4 + 0*8 + 1*16 + 0*32 = 19
3.6. Translation
The translation of braille features into their equivalent textual representation is achieved by
building a simple lookup table that holds a dictionary of features and their equivalent textual
representation. The translation of English braille, that holds few numbers of characters, is a
simple one-to-one-matching. But, Amharic has a large number of characters. Therefore, a
single textual character can be represented with either one or more braille cell and hence, one
or more features. This fact produces a huge challenge in building a lookup table for Amharic.
48
For this particular work, we have built a lookup table for 281 characters, 22 punctuation
marks and 20 numerals.
First, the extracted features are read and saved to a text file and an algorithm is designed that
matches every value in the text file against its corresponding Amharic textual representation
in the lookup table.
49
CHAPTER FOUR
EXPERIMENTATION
Braille is the preferred means of writing for visually impaired people that allows them the
ability to share their ideas to the world and also provide them access to what the world has in
braille written form. Digitizing braille can facilitate two way communications between vision
and visually impaired society. It can also enable to preserve literatures found in braille form
through electronic means which is a space saving and better secured option. However,
digitization of braille is not a simple process. Several countries have attempted to develop a
braille recognition system for their own languages. In Ethiopia, there were efforts made by
Teshome[43], Ibrahime[18], Miftah[32], Shumet[41] and Berihu[9] to develop braille
recognizer for local languages such as Amharic and Tigrigna. As a continuation to this efforts
which focused on single sided braille only, this study explored techniques that can allow
recognition of both single-sided and double-sided braille documents at once.
First, both single and double sided braille documents are collected and scanned using flatbed
scanner at a resolution of 200 dpi by applying the appropriate setting for light and contrast.
The digital image is preprocessed to reduce the noise and scale the image into relevant
regions only. Based on the orientation of colors (Black and White) in the dots the separation
of dots into recto and verso is performed. The important regions of the dots are segmented
and based on the segmentation result the features are extracted and passed to the translator
that matches features to their equivalent visual Amharic characters in the lookup table. This
chapter provides a detailed performance explanation of the main activities of the Amharic
Braille Recognition developed in this work.
4.1. Data collection
The data set used in this work is prepared from braille documents collected from Meserach
Center. Braille documents can be prepared with any color. The color of document used in this
work is mainly white and light yellow. This is because they are the most widely used and
available ones. The braille documents are standard 11 inches by 11.5 inches size. Braille can
be written single side or double-side. The focus of this work is the recognition of double-
sided Amharic braille document. But, the system is designed to work for single-sided braille
recognition as well. Therefore, the documents collected include both double-sided and single-
sided braille which are made of paper sheet embossed with the help of an embosser.
50
The detail of the sample braille documents collected for the purpose of experimentation is
given in table 4.1.
Dataset property Description
Number of Braille sheets
single-sided, 6
double-sided 30
Braille Type Grade I
Total number of characters 23,142
Average number of characters per sheet
single-sided, 387
double-sided 694
Color type White, Yellow
Resolution (horizontal/vertical) 200 dpi
Image size 12.5mb
Digital Format BMP
Document size 11" x 11.5"
Table 4.1 Database created
4.2. Digitization / Image Acquisition
The input to any braille recognition system is a digital image. Digital braille image can be
achieved by scanning braille documents with a scanner or by simply taking a picture with a
digital camera. Both techniques can provide the expected result. But, using digital camera
requires adjusting light source and angle of camera which is most of the time imperfect and
creates a skewed kind of image with non-uniform light intensity and shadow size. This
requires additional preprocessing to improve the image. Therefore, in this work, we have
51
preferred to use scanner which is a cheap alternative that can produce an excellent digital
image with minimal effort.
The type of scanner used in this project is an HP flatbed scanner that produces digital image
of the braille document into windows bitmap (BMP) format. The scanning is performed with
200 dots per inch (DPI) resolution which is the recommended level to get a good quality
braille image for subsequent processes. Braille can be written with any colored paper sheet.
The most dominantly used ones are gray, yellow, green and white. When scanning braille
sheets with lighter colors, such as white and light yellow, the light released from the machine
on the document during the scanning will affect the bright region of the dots making only the
shadow region to appear in scanned image. These images could not enable identification of
recto and verso dots. To solve this problem, we experiment other options such as scanning in
gray scale rather than full color, reducing contrast and applying adaptive lighting system.
From experiment, we have identified that the use of adaptive lighting system enable
producing a better quality digital image for any colored braille document. The kinds of braille
sheet used in this experiment are white, yellow and gray. Sample digitized braille is presented
in figure 4.1.
(a) (b) (c)
Figure 4.1 Scanned braille image; (a) ordinary scanning (b) gray scale scanning (c) scanning
with adaptive lighting
Since, adaptive lighting system is used in this work, the scanned braille images are full color
(RGB) images. Therefore, they must be converted to gray scale image before any further
processing. To do this, the color space value in the size of the scanned image is check. If the
color space value is three, it means the image is full color and should be converted to gray
52
scale. The MATLAB code for converting an RGB image to gray scale is shown in figure 4.2
below.
image=imread('D:\dataset\dset1\clean_01.bmp'); % get image

[m,n,z]=size(image);%Compute Size of Image, z is the color Space Value
if z==3
img1=rgb2gray(img);
else
img1=img;
end
Figure 4.2 MATLAB code for converting full color image into gray scale
4.3. Thresholding
In OCR systems, binarization is the technique that enables to separate the foreground
(characters) from the background. In this research, binarization is performed with a
thresholding approach. There are different thresholding techniques such as, global and local.
Global thresholding is used in this work with the reason being the appropriate approach for
binarizing an image having uniform color intensity such as braille. In addition, global
thresholding is computationally less expensive.
The main problem in using thresholding is the selection of threshold value to binarize the
image. Threshold value can be set manually or can be calculated from the pixel value of the
image using different technique. However, using a manually set threshold value will not
separate the foreground (dot) and the background properly. Moreover, it is not possible to use
a constant threshold value for all kinds of braille images since the scanned braille images
differ in their intensity level.
In this work, an initial attempt was made to define a threshold value for the image using the
matlab image processing toolbox function multithresh. The multithresh(A, 2) returns two
threshold values computed from the input image „A‟ by using Otsu‟s method. The two
threshold values are used as an input argument for the function imquantize, that convert
image „A‟ into 3 (2+1) discrete levels. This approach identifies the dots from the background
with some noise and it is observed that the dots are seen bigger in size and highly connected
to each other. In addition, it is seen that for some images with unusual intensity level, this
kind of thresholding doesn‟t extract the objects from the background clearly. Therefore, a
53
better thresholding technique that can handle the above problem is adopted from [3] and
modified to improve result.
In Thresholding the image, the algorithm begins by calculating the average gray level of the
whole image. After the average value is calculated, it defines the maximum and minimum
values for each column of the image. Then from the maximum and minimum it computes the
values for „a‟ and „b‟; if the value of a given pixel of the image is greater than „a‟, then the
value is added to the variable „y‟ and if the value of a given pixel is less than „b‟, then the
value is added to the variable „x‟. Finally, „H‟ and „L‟ are calculated based on max_avg and
min_avg values. „H‟ and „L‟ are the thresholding values such that, if the value of a pixel is
greater than „H‟, then the pixel will be converted to white; or if the value of the pixel is less
than „L‟, it will be converted to black; otherwise it is converted to average value of the image.
The MATLAB implementation of the algorithm for thresholding is given in Figure 4.3.
%% Thresholding %img3 is filtered braille imahe

immean=mean(mean(img3)); % avg pixel value for the image
a=mean(max(img3)+immean)/2; b=mean(min(img3)+immean)/2;
[M N]=size(img3); %get size of the image
y=0; x=0;
for i=1:M
for j=1:N
if img3(i,j)> a ;
q=img3(i,j);
y=[y,q];
elseif img3(i,j)< b ;
w=img3(i,j);
x=[x,w];
end
end
end
max_avg=mean(y);
H=immean+(max_avg-immean)/3;
min_avg=mean(x);
L=immean - (immean - min_avg)/3; for i=1:M
for j=1:N
if img3(i,j)> (H)
img3(i,j)=255;
elseif img3(i,j) < (L)
img3(i,j)=0;
else
img3(i,j)=immean;
end
end
end
Figure 4.3 MATLAB code for thresholding
54
The algorithm is implemented on matlab and has produced good result with any kind of
braille image as compared to the built-in matlab algorithm.
The image in figure 4.4 presents a sample RGB image and its binarized representation.
(a) (b)
Figure 4.4 Sample thresholded image: (a) Scanned RGB Image (b) image a after thresholding
As mentioned above, braille document by their very nature can have their quality affected
due to repeated use, scratch and bend. And in double sided writing the process of embossing
itself will degrade the dots and the paper itself. These problems can affect the thresholding
performance.
4.4. Separation of dots for Double-Sided Braille
The major challenge in recognition of double-sided braille is the separation of dots of recto
and verso. The writing system is designed in such a way that the verso dots appear perfectly
between the recto dots. In a scanned braille image, the recto and verso dots are found
connected to each other. But, the relief is in that those dots have different color orientation.
Therefore, the technique used in this work exploits this behavior to separate the dots. The
algorithm is adopted from [3]. It works by filtering the image column-wise with a window of
8 by 1 pixels and loop through all columns in the image. A window size of eight is selected
because in the experiment it is observed that the average window height for recto dots and
verso dots is 8 and 7 pixels respectively. Therefore, a general window size of 8 pixels is used
to work on the dots of both sides. Part of the MATLAB implementation of the code can be
seen in figure 4.5.
55
%separate recto and verso...
Newval=uint8([255;255;255;255;0;0;0;0]);
windowHeight = 8; windowWidth =1;
imagej=ones(size(img3));
RectoImg=zeros(imageHeight,imageWidth,dim);
VersoImg=zeros(imageHeight,imageWidth,dim);
for i = 1:imageWidth - windowWidth + 1
j=1;
while j < imageHeight - windowHeight + 1
if imagej(j,i)==255
j=j+8;
else
m=j;
j=j+1;
window = img3(m:m + windowHeight - 1, i:i + windowWidth - 1);
if (window(1, 1) ~= (255||immean) && window(2, 1) == 255 ...
&& window(3, 1) == 255 && window(7, 1) == 0 && ...
window(8, 1) == 0);
RectoImg(m:m + windowHeight - 1,i:i + windowWidth - 1) = ...
immultiply(window , newval);
imagej(m:m + windowHeight - 1,i:i + windowWidth - 1) = 255;
elseif (window(1, 1) ~= immean && window(2, 1) == 0 && ...
window(3, 1) == 0 && window(6, 1) == 255 && ...
window(7, 1) == 255 && window(8, 1) ~= immean);
VersoImg(m:m + windowHeight - 1,i:i + windowWidth - 1) = ...
immultiply(window, flipdim(newval, 1));
imagej(m:m + windowHeight - 1,i:i + windowWidth - 1) = 255;
end
end
end
end
Figure 4.5 Part of MATLAB code for front and back side dot separation
The algorithm starts by creating two separate arrays to hold the recto and verso dots. For
every window, the filter checks if the upper pixels are white and the lower are black. If this
check results true, it registers it to the recto array. But if the upper pixels are black and the
lower pixels are white, the result is written on the verso array.
Sample double-sided braille image and the result of separating the recto and verso dots can be
seen in figure 4.6. The filtering is made in such a way that it will not check same region
again. This technique had reduced computational time and number of false recognitions.
56
(a) (b)
(c) (d)
Figure 4.6 Recto and Verso dots separation; (a) scanned RGB image, (b) Gray Scale image
showing the recto and verso dot position (Red- Recto, Blue - Verso), (c) recto dot regions
after separation and (d) verso dot regions after separation.
This algorithm is implemented on MATLAB and had produced the desired result on any kind
of braille documents. Though it is designed to separate the recto and verso dots, it can also
help to detect regions of dots in single-sided writing as well. At the end of this process, we
have two binary images holding the recto and verso dot regions separately. But, it is to be
recalled that as the level of noise increase, defects start to be noticed. Some of the defects
identified are:-
 A dot region in the threshold image starts to lose form. As a result, it is difficult to
find the dot regions in this process.
57
 Dots will be highly connected and due to imperfection of the thresholing operation, a
dot region may be detected with fewer pixels, none dot region may be detected as a
dot region and a dot region may be detected with many pixels.
These and other problems highly affect subsequent operation of braille cell formulation.
Therefore, in order to reduce the effect of this problem, morphological operations are
performed that can improve the shape and size of detected regions. On average, after
detection, a dot area has 3 pixels height and 4 pixels width. But, some dots may be identified
as two or more separate regions having 1 pixel width. Therefore, to keep these regions as
valid ones, we have implemented an algorithm that performs a neighborhood search and
connect such separate regions. In addition, it is observed that some of the detected dot regions
do not form perfect rectangles. This will diminish the vertical space between lines of dot in a
cell. Therefore, another morphological operation is performed that improves the shape of dot
regions detected. Sample binary images of braille before and after morphological operation
are performed can be seen in figure 4.7 below.
Figure 4.7 Binary braille dote regions before and after morphological improvement
4.5. Braille Cell formulation

In order to group the braille dots into cell; the technique used in this research is the
construction of grid. Grids are constructed by using the horizontal and vertical projection
profile of perfectly positioned braille dot. Therefore, in this work, only perfectly positioned
braille dots are used in building the grid. To do this, an algorithm that performs automatic
area analysis and define the average area „K‟ for the dots and filters out dots that are less than
or greater than 2±k pixel counts is implemented. The lower and higher bounds are defined
with experiment to work for any noise level braille documents.
After excluding improperly sized dots, which bias the result of horizontal and vertical
projection, a few numbers of dots in the image remain. With the properly sized and placed
dots, the horizontal and vertical projection is developed and produces the expected result.
58
Part of the MATLAB implementation of this algorithm is shown in figure 4.8 below.
% Perform area analysis with 4 connectivity

imgbw = bwconncomp(img4bw, 4);
label = labelmatrix(imgbw);
bwarea = regionprops(imgbw, 'Area');
K=round(mean(bwarea.Area);
% Remove big pixels

rimage2 = ismember(L, find([bwarea.Area] <= K + 2 ));
% Remove small pixels

rimage2 = ismember(L, find([bwarea.Area] >= K - 2 ));
%% Horizontal Projection
[imageheight, imagewidth] = size(rimage2);
hp = zeros(,1);
for i = 1: imageheight
for j = 2: imagewidth
if rimage2(i,j)~=rimage2(i,j-1)
hp(i) = hp(i)+1;
end
end
end
%% Vertical projection front

vp = zeros(1, imagewidth);
for j = 1: imagewidth
for i = 2: imageheight
if rimage2(i,j)~=rimage2(i-1,j)
vp(j) = vp(j)+1;
end
end
end
Figure 4.8 Part of the MATLAB code for building grid
The coordinate values to build the grid line are derived from the position of the center points
of the connected components in the horizontal and vertical projection results. From the center
points in the vertical projection, we build vertical lines connecting all the points in „y‟ axis.
From center points in the horizontal projection, we build horizontal lines connecting all the
points in „x‟ axis.
Sample braille grid image constructed with all the dots in the image and after removing
improperly sized dots can be seen in figure 4.9 below.
59
Figure 4.9 Results of grid construction; (Top) with all dots, (Bottom) with selected dots
This way we can build grids line over the points only where there is a dot. But, it is a must to
find the possible dot regions as well although they do not have a dot at the time. Therefore, in
order to find those missed points, we have designed a grid refining algorithm that analyzes
the horizontal and vertical spacing between dots in a cell and between cells themselves. This
algorithm builds grid line where they should exist. The output of this operation is shown in
figure 4.10 below.
60
Figure 4.10 Results of grid refinement
4.6. Amharic Braille feature extraction
The extraction of braille feature is performed by checking the existence of dot on or around
the intersection points of the grid line. Six intersection points which are closer to each other
make up feature for single cell. This way the feature of every cell will be extracted and saved
in an array. The array holding the features is an array of zeros and ones. Zeros indicate that
there is no dot in the specified position while ones indicate the existence of dot. The feature
for a single cell is composed of six binaries. However, for simplicity the six binary values
will be converted to decimal code representation.
61
The MATLAB implementation of the algorithm is given in figure 4.11 below.
%% save coordinates to array for recto...
pp1=[];
cc1 = 1;
for nt = [1:num12]
X = centroids12(nt:nt,2);
for cn = [1:num22]
Y=centroids22(cn:cn,1);
pp1(cc1,1) = X;
pp1(cc1,2) = Y;
cc1 = cc1 + 1;
end
end
%% extract features
A=[];
si=size(pp1);
n=1;
segi=zeros(size(rimage));
sizei=size(centroids12,1);
sizej=size(centroids22,1);
for o =(1:3:sizei) % loop through the first line of every cell

c=o;
for w =(1:sizej)
j=centroids22(w:w,1);
x=round(j);
for y =(c:c+2)
i = centroids12(y:y,2);
y=round(i);
if rimage(y,x)==1 || rimage(y,x-1)==1 || rimage(y,x+1)==1
B=1;
A=[A;B];
else
B=0;
A=[A;B];
end
end
end
end
%% convert extracted binary features to decimal code representation
d=[]; % Array to hold the decimal codes

for i=1:6:(size(A,1)-5)
w=A(i:i+5);
decval=w(1)*1 + w(2)*2 + w(3)*4 + w(4)*8+ w(5)*16 + w(6)*32;
d=[d; decval]; % append to d
txtout = fopen('result_01.txt', 'w'); % Save to text file
fprintf(txtout, '%d\n', d);
end
Figure 4.11 MATLAB code for feature extraction
62
4.7. Translation
The final step in the designed system is translation. Here we convert the feature, which is a
decimal code representation for every cell, into its corresponding Amharic letters to get its
translation. Previous works have used classification techniques such as SVM, neural network
and j48. However, braille translation is a one-to-one matching. Therefore, building a lookup
table and matching features against it will be sufficient.
To build the translator, it is first required to understand the context of the language. In
Amharic braille writing a single cell can represent a single character, part of a character, a
numeral or punctuation mark. All the 34 core characters have a single cell representation
while the six variants of the 34 core characters, which make a total of 204 (34*6), are
represented by two cells. The numerals are represented with two cells while punctuations are
represented with either two or three cells. Sample characters and their normalized
representation can be seen from table 4.2.
First cell Second cell Third cell Lookup

characters table
Decim Decim Decim
represent
1 2 3 4 5 6 al 1 2 3 4 5 6 al 1 2 3 4 5 6 al
ation
Code Code Code
ሀ 1 1 0 0 1 0 19 0 1 0 0 0 1 34 0 0 0 0 0 0 0 1
ሁ 1 1 0 0 1 0 19 1 0 1 0 0 1 37 0 0 0 0 0 0 0 2
ሂ 1 1 0 0 1 0 19 0 1 0 1 0 0 10 0 0 0 0 0 0 0 3
ሃ 1 1 0 0 1 0 19 1 0 0 0 0 0 1 0 0 0 0 0 0 0 4
ሄ 1 1 0 0 1 0 19 1 0 0 0 1 0 17 0 0 0 0 0 0 0 5
ህ 1 1 0 0 1 0 19 0 0 0 0 0 0 0 0 0 0 0 0 0 0 6
ሆ 1 1 0 0 1 0 19 1 0 1 0 1 0 21 0 0 0 0 0 0 0 7
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
1 0 0 1 1 1 1 60 1 0 0 0 0 0 1 0 0 0 0 0 0 0 283
፩ 1 1 1 1 1 1 63 1 0 0 0 0 0 1 0 0 0 0 0 0 0 294
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
፡ 0 0 0 0 0 1 32 0 0 1 0 0 0 4 0 0 0 0 0 0 0 313
… 0 0 1 0 0 0 4 0 0 1 0 0 0 4 0 0 1 0 0 0 4 305
Table 4.2 Sample Amharic features and their normalized representation
As can be seen in lookup table Amharic characters have vowel and consonant combinations.
The decimal codes for the vowels are 34, 37, 10, 1, 17 and 21 which represent the 1 st, 2nd,
63
3rd,4th, 5th and 7th sounds of a 7 order character. Therefore, any number followed by those
decimal values is a consonant for a character represented with two cells. Any decimal code
followed by preceded by 60 is a Latin numeral while any decimal value preceded by 63 is a
geez numeral. Every punctuation mark has its own decimal representation which could be in
two or three cell. Every vowel and consonant combination, every numeral and every
punctuation mark is assigned a single unique numeral for uniformity. Finally a lookup table is
built with python that contains the character and their respective numeral representation. The
translation code will group numbers and match the cell representation of the features
extracted and saved in the text file against their corresponding Amharic characters in the
lookup table. For example if converting a binary future value to decimal results in number 19
followed by 37 it is assigned the unique number 2 in lookup table and the translation is
Amharic Character "ሁ". The complete lookup table can be found in appendix I. A single
word along with its features and translation steps can be seen in figure 4.12 below.
-Grid construction for feature extruction
- Extracted features (dots represented by 4 pixels)
- Features of dot in binary representation

101111 100000 011100 010001 101110 100000 110000 011110 100000
100001 010001 010111 100000 111000 010011
- Fatures of dot in decimal representation with consonant voul grouping
61,1 14,34 29,1 3 30,1 33,34 58,1 7 50
- Lookup table representation for consonant voul groupings
158 29 95 69 74 78 130 13 319
- Translation agrainst lookup table
ያ ሰ ና ብ ታ ቸ ዋ ል ፡፡
Figure 4.12 Steps of translation
64
4.8. Performance Evaluation
The performance of the double-sided Amharic braille recognizer was tested using three
separate datasets classified based on noise level; clean, medium noisy and high noise images.
Each dataset contains images of 10 double-sided and 2 single-sided Amharic braille
documents. In order to report the performance of the system on single-sided and double sided
separately, each dataset is gain classified into two; the first group contains 2 single-sided
braille images in each noise level and the second dataset contains 10 double-sided braille
images in each noise level.
The results achieved are detailed in table 4.4 and 4.5 below. First, the performance of the
system is measured after segmentation on dots level. The total number of potential dots is
defined by manually counting the total dots region in every of the dataset which includes both
the front side and back side dot. There are two kinds of errors that affect the performance.
These are dots added (false positive) and dots missed (false negative). Dots added is the total
number of none dot areas recognized as dot while the dots missed is a true dot region
recognized as a background. Segmentation accuracy is measured as a percentage of correct
recognition against total recognition.
Segmentation
Total number of Accuracy
Potential dot (𝐴 − 𝐵 − 𝐶)
Test Set areas in the Dots Dots 𝐴
braille sheet added missed
(A) (B) (C) (%)
Test set 1 (clean) 14,472 0 0 100.00

Test set 2
(medium level noise) 13,932 28 11 99.72
Test set 3 (high noise) 14,112 438 107 96.14
Average 42,516 466 118 98.63
Table 4.4 Performance results for single-sided segmentation
65
Total number of
Segmentation
Potential dot
Accuracy
Test Set areas in the Dots Dots
(𝐴 − 𝐵 − 𝐶)
braille sheet added missed 𝐴
(A) (B) (C) (%)
Test set 1 (clean) 68,040 0 0 100.00
Test set 2 (medium
level noise) 69,600 174 210 99.45
Test set 3 (high
noise) 68,718 12054 1971 79.59
Average 206,358 12228 2181 93.02
Table 4.5 Performance results for double-sided segmentation
As can be seen in table 4.4 and 4.5 the performance of the system on both single-sided and
double-sided clean documents is 100%. However, the performance reduces as the level of
noise in braille documents increases. It is also observed that the average performance of the
system on single-sided is slightly better than that of the average performance registered on
double-sided braille documents. This is because single-sided documents are less noisy than
that of double-sided documents. But, the main reason lays in the fact that the number of dots
in double-sided documents is almost twice that of dots in single-sided. This makes the
number of connected dots in double-sided braille to be very high as compared to single-sided
braille dots which are rarely connected.
As can be seen from the table an average accuracy of 98.63% and 93.02% are registered on
segmentations of single-sided and double sided braille documents respectively. Sample
braille image and segmentation performance of the system on the three noise levels can be
seen in figure 4.13 below.
66
(a) (d) (g)
(b) (e) (h)
(c) (f) (i)
Figure 4.13 Segmentation results (a) Clean braille image (b) recto dots extracted from a (c)
verso dots extracted from a (d) Average noisy braille image (e) recto dots extracted from d
(f) verso dots extracted from d (g) High noisy braille image (h) recto dots extracted from g
(i) verso dots extracted from g
It is also measured the accuracy of the system at the level of correctly translated characters.
At this level the overall performance of the system is tested. The results on single-sided
braille and double-sided braille images are detailed in table 4.6. and 4.7 respectively.
67
Total Incorrectly
number of recognized Segmentation
Test Set
characters Characters Accuracy
(A) (B) (%)
Test set 1
2,412 0 100
(clean)
Test set 2
2,322 6 99.74
(medium level noise)
Test set 3
2,352 258 89.03
(High level noise)
Average 7,086 264 96.27
Table 4.6 Overall performance achieved on single sided recognition

Total Incorrectly
number of recognized Segmentation
Test set characters Characters Accuracy
(A) (B)
(%)
Test set 1
11,340 0 100.00
(clean)
Test set 2
11,600 57 99.51
(medium level noise)
Test set 3
11,453 4015 64.94
(High level noise)
Average 34,393 4072 88.16
Table 4.7 Overall performance achieved on double-sided recognition
As can be seen from table 4.6 and 4.7 above, the system had achieved an overall accuracy of
100% on clean braille images of both single-sided and double-sided braille documents. But,
the overall performance of the system fails as noise increase, the same way as the test in
segmentation stage. This is because the segmentation output highly affects the grid
construction process. Since feature extraction depends on correctly positioned grids, miss
placing a single grid can highly affect the accuracy in extracting features. But, it is also to be
noted that correctly placed grids have the effect of increasing accuracy by cutting out
incorrectly segmented dots. The overall accuracy achieved is 96.27% for single-sided and
88.16% for double-sided braille documents.
68
4.9. Discussion and Challenges
The system designed has registered a remarkable result on both clean and average single-
sided and double-sided braille documents. The reason for achieving such result is due to
effectiveness in the thresholding techniques and dot extraction mechanism used in this work,
which also helps to exclude noise from the image. It is also worth mention that the grid
construction and refining technique introduced in this work has also contributed a significant
role in extracting features correctly. It is difficult to compare performance of the current
system against previous works for single-sided braille recognition. But, since double-sided
braille recognition is more difficult, due to the addition in noise and number of feature to be
extracted, it is sound to say that the results registered are commendable as compared to
previous works for single-sided. Sample braille image and translation result can be found on
Appendix II.
It is also to be noted that, the performance of the current system is highly affected on high
level noise documents. The errors encountered on high noisy documents are due to the
quality of the braille document and the performance of the thresholding operation which in
turn affects subsequent operations. In this work we have tried to adopt a thresholding
algorithm that dynamically calculates the threshold values for any kind of image. But, the
performance of the thresholding operation is slightly affected with noisy and degradation.
When dots are degraded the thresholding algorithms fails to build them with the proper
orientation of black and white. Some of the problems identified during thresholding can be
seen in figure 4.14 below. Red boxes represent segmentation errors observed.
Figure 4.14 Problems identified in thresholding
69
These slight defects are seen to have huge impact on the performance of the segmentation
and feature extraction processes.
Another thing worth mentioning is, skew detection and correction. Due to time limitation,
this system is not designed to detect and correct skew. As a result it will not work properly
for skewed images with a skew angle less that -0.5 and greater than +0.5. However, we have
tried to manually correct skew images and check if the recognizer performs properly. The test
had shown that the recognizer performed very well.
70
CHAPTER FIVE
CONCLUSION AND RECOMMENDATION
5.1. Conclusion
For visually impaired people, lack of vision has been an obstacle to access printed contents as
well as write their ideas on paper. In the long effort of solving this important problem Luis
Braille invented a writing system called brialle, which is named after him. Braille is an
efficient means of writing and communication for the blind. It has been widely adopted by
different counties for use with their own languages. Amharic is one of the languages that
Adopted Braille.
The invention of Braille has solved the communication through written form problem
between visually impaired and vision people. However, the use of Braille for communication
with vision society requires vision people to understand braille. OBR is a technology with the
answer for this problem. It is a system that converts hardcoded braille documents into
machine readable and editable text that vision people can understand. In addition, OBR has
allowed preservation of Braille works and it is also an easy way to copy Braille.
In this work an Amharic Braille Recognition System is developed. The developed system is
able to recognize both single-sided and double-sided Amharic braille documents. Amharic,
unlike English has very large number of characters and hence very complex braille
representation. In order to accommodate the large number of characters, a single character
can be represented with up to 3 cells. Understanding this property is important in building
braille recognition system for the language. In this work, different techniques for the
recognition process adopted from previous works with some modification and also designed
new algorithms for improving the performance of the system. Image noise filtering
techniques that were recommended by previous researchers on Braille images have been
tested and used. A dynamic thresholding technique is adopted that computes different
threshold values for different images automatically. The main problem, separation of front
and back side dots, have been solved by using a sliding window that filters the image to
detect dot areas and separate the dots into front and back side dots.
An improved grid construction technique, that automatically define and sort dot size and uses
only perfectly sized dots, have been designed to build a grid over the image. We have also
71
introduced a technique that refines the grid based on the horizontal and vertical spacing
property between dots and between cells. This technique works by taking a sample from
perfectly constructed grid lines. It has the advantage of flexibility in automatically defining
the inter cell and intra cell distance rather than calculating the dot size and distance which
previous researches have been doing.
The segmentation technique and most importantly the grid construction process have enabled
the detection of dots with high performance. A prototype has been designed for the system
and its performance have been tested against three sets of braile images; clean, average noisy
and high noisy. The designed system has registered higher results for clean and average noisy
braille documents while a small error is registered in recognizing high noisy Braille
documents. As a result, an average accuracy of 96.07% and 86.16% has been achieved for
single-sided and double sided braille documents respectively.
The commendable results achieved are due to the effectiveness in the thresholding,
segmentation and feature extraction techniques. The small error observed on the noisy
documents is due to the defects on the Braille document that alter the shape of the dots. Such
problems have an impact on the thresholding performance and this in turn affects accuracy in
segmentation of dots.
The performance of any character recognition system is highly dependent on the quality of
image it process. As a result, a lot of consideration should be done in selecting the
preprocessing techniques to use. The preprocessing technique should be able to provide the
image as per the requirement for the segmentation. The major challenge unique double-sided
recognition is the separation of verso and recto dots. In order to achieve good result in
separation, the algorithm should focus on the unique property of the dots of images of recto
against verso. Such uniqueness can be found in the orientation of bright and shadow regions.
In this work, exploiting this nature of braille image has enabled separation of recto and verso
dots.
This work has added to the search for comprehensive Amharic Braille recognition system by
working on double-sided braille recognition that has not been locally attempted so far. The
performance of the system on single-sided documents is also better compared to previous
works for Amharic braille.
72
5.2. Recommendation
This study has attempted to adopt and design an Amharic OBR that recognizes both single-
sided and double-sided Braille documents. Based on the findings from the study, the
following recommendations are forwarded for future researches in the area,
 The proposed system detects dots based on the bright and dark region orientation. A
dot area to be detected should have both properties. But, due to quality of the Braille
document, in the threshold image either of the two regions might not appear.
Therefore, this dot will be considered as a background. Future works can focus on
recovering such dots to improve the performance of the system.
 In this work a vertical scan is used to checks the existence of bright and shadow areas.
In case of skewed images, correcting the skew angle will make the bright and shadow
area to take a horizontal position rather than vertical orientation. As a result this
technique will fail to detect most of the dots. Therefore, future works should search
for recognition techniques of skewed braille images.
 In this work separation of recto and verso dots is dependent on the black and white
regions of dots formed during scanning. However, braille documents are mostly
degraded and it is difficult to get a clean image with clear separation for the black and
white regions. Therefore, future works can explore techniques that can crop out dots
by matching features with model designed using a learning system which is
preferable.
 Braille documents by their very nature are prone to physical damaged. As a result
almost all braille images are noisy and noise decreases the performance of a system.
But, the overall error registered in this system is very small. Therefore, integrating
post recognition process that corrects wrongly recognized words by checking them
against their closest match in a dictionary can improve performance of the system.
 In this work, no effort was made in keeping the format of the Braille after translation.
However, keeping the format might be important in such cases where the content of
the Braille is a poem. Therefore, future works can integrate techniques for keeping the
format of Braille after translation.
73
Reference
[1] A. Papandreou and B. Gatos, 'A novel skew detection technique based on vertical
projections', Document Analysis and Recognition (ICDAR), International Conference
on (pp. 384-388), IEEE, 2015.
[2] A. Abdelmonem, B. El-Hoseiny, C. Ali, D. Emara, E. Hafez and F. Gamal, 'Dynamic
Optical Braille Recognition (OBR) System', IPCV, pp. 779-786, 2009.
[3] A. Al-Salman, Y. AlOhli, M. AlKanhal and A. AlRajih, 'An Arabic Optical Braille
Recognition System', Proceedings of the International Conference in Information and
Communication Technology & Accessibility (ICTA‟07), pp. 81-86, 2007.
[4] A. Antonacopoulos and D. Bridson, 'A robust Braille recognition system', In Document
Analysis Systems VI, pp. 533-545, 2004.
[5] A. Cheung, M. Bennamoun and N. Bergmann, 'An Arabic optical character recognition
system using recognition-based segmentation', Pattern Recognition, vol. 34, no. 2, pp.
215-233, 2001.
[6] A. Mousa, H. Hiary, R. Alomari and L. Alnemer, 'Smart Braille System Recognizer',
IJCSI International Journal of Computer Science Issues, vol. 10, no. 6, 2013.
[7] Al-Saleh, 'Dot Detection of Braille Images Using A Mixture of Beta Distributions',
Journal of Computer Science, vol. 7, no. 11, pp. 1749-1759, 2011.
[8] B. Sankur, M. Sezgin 'Survey over image thresholding techniques and quantitative
performance evaluation', Journal of Electronic Imaging , vol. 13, no. 1, p. 146-165,
2004.
[9] B. Yohannes, 'Recognition of Noisy Ethiopic Braille Documents', MSc. Thesis, Addis
Ababa University, School of Information Science, Addis Ababa, Ethiopia 2013.
[10] C. Kothari, Research methodology. New Delhi: New Age International (P) Ltd., 2004.
[11] C. M. Ng, V. Ng and Y. Lau, 'Statistical Template Matching for Translation of Braille',
In Proceedings of the Spring Conference on Computer Graphics (SCCC), pp. 197-200,
1999.
[12] C. Mackenzie, World braille usage. [Paris]: UNESCO, 1954.
[13] C. Ng, V. Ng and Y. Lau, 'Regular Feature Extraction for Recognition of Braille',
Computational Intelligence and Multimedia Applications, ICCIMA „99. pp. 302-306,
1999.
[14] Csa.gov.et, 'Census 2007', 2015. [Online]. Available:
http://www.csa.gov.et/index.php/2013-02-20-14-51-51/2013-04-01-11-53-00/census-
2007. [Accessed: 03- Jun- 2015].
[15] D. Bradley and G. Roth, 'Adaptive Thresholding using the Integral Image', Journal of
Graphics, GPU, and Game Tools, vol. 12, no. 2, pp. 13-21, 2007.
74
[16] D. Saad Al-Shamma and S. Fathi, ”Arabic Braille Recognition and Transcription into
Text and Voice”, 2010 5th Cairo International Biomedical Engineering Conference
Cairo, Egypt, December 16-18, 2010.Arabic to text and voice
[17] Data.worldbank.org, 'Ethiopia | Data', 2015. [Online]. Available:
http://data.worldbank.org/country/ethiopia. [Accessed: 03- Jun- 2015].
[18] E. Chekol, 'Recognition of Amharic Braille Documents', MSc. Thesis, Addis Ababa
University, Department of Information Science, 2010.
[19] E. Davies, Computer and machine vision. Waltham, Mass.: Elsevier, 2012.
[20] Encyclopedia Britannica, 'Amharic language', 2013. [Online]. Available:
http://www.britannica.com/EBchecked/topic/20500/Amharic-language. [Accessed: 03-
Jun- 2015].
[21] Ethiopians.com, 'Ethiopian Writing System - Baye Yimam', 2015. [Online]. Available:
http://www.ethiopians.com/bayeyima.html. [Accessed: 03- Jun- 2015].
[22] G. Morgavi and M. Mauro, 'A neural network hybrid model for an optical braille
recognitor', International Conference on Signal, Speech and Image Processing
(ICOSSIP), 2002.
[23] H. M. Taha,. Robust braille recognition system using image preprocessing and feature
extraction algorithms (Doctoral dissertation, Universiti Tun Hussein Onn Malaysia)
2014.
[24] J. Abualkishik, M. Abdallah and O. Khairuddin, 'Quranic Braille System', International
Journal of Humanities and Social Sciences, pp. 313-319, 2009.
[25] J. Bhattacharya and S. Majumder, 'braille character recognition using a generalized
feature vector', Computer Networks and Intelligent Computing Springer Berlin
Heidelberg, pp. 171-180, 2011.
[26] J. Mennens, L. Van Tichelen, G. Francois and J. Engelen, 'Optical recognition of
Braille writing using standard equipment', IEEE Trans. Rehab. Eng., vol. 2, no. 4, pp.
207-212, 1994.
[27] J. Mennens, L. Van Tichelen, G. Francois and J. Engelen, 'Optical recognition of
Braille writing', IEEE Trans. Rehab. Eng., vol. 2, no. 4, pp. 207-212, 1993.
[28] J. Sadri, C.Y.Suen and T.B.Bui, 'Application of support Vector Machine for
Recognition of Hand Written Arabic/Persian Digits', Center for Pattern Recognition
and Machine Intelligence, 2004. MVIP vol. , 2004
[29] L. Shapiro and C. Stockman G., Computer Vision. Prentice Hall, 2011.
[30] L.Wong, W., Abdulla, & S. Hussmann, A software algorithm prototype for optical
recognition of embossed Braille. In Pattern Recognition, 2004. ICPR 2004.
Proceedings of the 17th International Conference on (Vol. 2, pp. 586-589). IEEE,
2004.
75
[31] M. Cheriet, N. Kharma, L. Liu C. and C. Suen, Character recognition systems: a guide
for students and practitioners. Hoboken, N.J.: Wiley-Interscience, 2007.
[32] M. Hassen, 'Recognition of Amharic Braille Using Direction Field', MSc. Thesis,
Addis Ababa University, Department of Computer Science, 2011.
[33] M. Meshesha, 'A Generalized Approach to Character Recognition of Amharic Texts',
MSc. Thesis, Addis Ababa University, School of Information Studies for Africa, 2000.
[34] M. Wanas N, A. Said D, H. Hegazy N and M. Darwish N, 'A study of local and global
thresholding techniques in text categorization', In Proceedings of the fifth Australasian
conference on Data mining and analystics, vol. 61, pp. 91-101 Australian Computer
Society, Inc., 2006.
[35] M. Yousefi, M. Famouri, B. Nasihatkon, Z. Azimifar and P. Fieguth, 'A robust
probabilistic Braille recognition system', International Journal on Document Analysis
and Recognition (IJDAR), vol. 15, no. 3, pp. 253-266, 2011.
[36] Matlab, Mathworks, Matlab Image processing toolbox, 2013
[37] N. Falcon, C. Travieso, J. Alonso and M. Ferrer, 'Image processing techniques for
braille writing recognition', Computer Aided Systems Theory–EUROCAST 2005
Springer Berlin Heidelberg, pp. 379-385, 2015.
[38] R. Gonzalez, R. Woods and S. Eddins, Digital Image processing using MATLAB.
Upper Saddle River, N.J.: Pearson Prentice Hall, 2004.
[39] R. Jain, R. Kasturi and B. Schunck, Machine vision. New York: McGraw-Hill, 1995.
[40] S. T and V. Udayashankara, 'A Review on Software Algorithms for Optical

Recognition of Embossed Braille Characters', International Journal of Computer
Applications, vol. 81, no. 3, pp. 25-35, 2013.
[41] S. Tadesse, 'Feature Extraction And Classification Schemes For Enhancing Amharic
Braille Recognition System', Msc. Thesis, Addis Ababa University, School of
Information Science, 2011.
[42] Simpson, The rules of unified English Braille, 2nd ed. International Council on English
Braille(IECB), 2013.
[43] T. Alemu, 'Recognition of Amharic Braille', MSc. Thesis, Addis Ababa University,
Department of Information Science, 2009.
[44] T. Shreekanth, V. Udayashankara,. An Algorithmic Approach for Double Sided Braille
Dot Recognition Using Image Processing Techniques. In International Journal of
Image Processing and Visual Communication ISSN (Vol. 2, issue 4). 2014
[45] T. Pang-Ning, S. Michael and K. Vipin, Introduction to data mining. Addison Wesley,
2006.
76
[46] UNESCO, 'World Braille Usage', National Library Service for the Blind and Physically
Handicapped Library of Congress, Perkins International Council on English Braille,
Washington, D.C, 2013.
[47] UNESCO, 'World Braille Usage', National Library Service for the Blind and Physically
Handicapped Library of Congress, Perkins International Council on English Braille,
Washington, D.C., 1990.
[48] V. Udayashankara, 'An Application of Eight Connectivity based Two-pass Connected-
Component Labelling Algorithm For Double Sided Braille Dot Recognition',
International Journal of Image Processing (IJIP), 8(5), 294., vol. 8, no. 5, pp. 220 - 396,
2014.
[49] Who.int, 'WHO | Blindness', 2015. [Online]. Available At:
http://www.who.int/topics/blindness/en/ [Accessed: 05- Sep- 2014].
[50] X. Hermida, A. Rodríguez and F. Rodríguez, 'A Braille OCR for Blind People',
Proceedings of ICSPAT-96. Boston (USA). 1996.
[51] x!T×ùÃ xYnSW‰N B/@‰êE ¥HbR 12¾ z#R S‰ xSfÉ¸ ÷¸t&

ytÌÌmW yx¥R¾ BÊL xššY ÷¸t&ÝÝ yx¥R¾ BÊL xššY ÷¸t& zgÆÝÝ
¬HúS 19 xÂ 2; qN 1995 ›.M. lt-‰W xgR xqF g#Æx@ yqrbÝÝ
[52] nByL;#L ×¦NS zmn BR¦N xÄ!S xbÆ BR¦NÂ s§M ¥t¸Ã b@T 1954 ÝÝ
[53] Omniglot.com,. 'Amharic Alphabet, Pronunciation And Language'. N.p., 2015. Web.
20 Oct. 2015.
77
Appendix
I. List of the seven braille orders
Order1
Order 2
Order 3
Order 4
Order 5
Order 6
Order 7
Table I List of the seven braille orders
78
II. Complete list of Amharic Alphabet
Amharic Characters
order Labialized
st nd rd
1 2 3 4th 5th 6th 7th
ሀ ሁ ሂ ሃ ሄ ህ ሆ
ሇ ለ ሉ ሊ ላ ሌ ል ሎ
ሏ ሐ ሑ ሒ ሓ ሔ ሕ ሖ
ሗ መ ሙ ሚ ማ ሜ ም ሞ
ሟ ሠ ሡ ሢ ሣ ሤ ሥ ሦ
ሧ ረ ሩ ሪ ራ ሬ ር ሮ
ሯ ሰ ሱ ሲ ሳ ሴ ስ ሶ
ሷ ሸ ሹ ሺ ሻ ሼ ሽ ሾ
ሿ ቀ ቁ ቂ ቃ ቄ ቅ ቆ ቇ ቈ ቉ ቊ
በ ቡ ቢ ባ ቤ ብ ቦ ቧ
ቨ ቩ ቪ ቫ ቬ ቭ ቮ ቯ
ተ ቱ ቲ ታ ቴ ት ቶ ቷ
ቸ ቹ ቺ ቻ ቼ ች ቾ ቿ
ኀ ኁ ኂ ኃ ኄ ኅ ኆ ኇ ኈ ኉ ኊ ኋ
ኌ ኍ ኎ ኏ ነ ኑ ኒ ና
ኔ ን ኖ ኗ ኘ ኙ ኚ ኛ
ኜ ኝ ኞ ኟ አ ኡ ኢ ኣ
ኤ እ ኦ ኧ ከ ኩ ኪ ካ ኬ ክ ኮ ኯ
ኰ ኱ ኲ ኳ ኴ ኵ ኶
ወ ዉ ዊ ዋ ዌ ው ዎ
ዏ ዐ ዑ ዒ ዓ ዔ ዕ
ዖ ዗ ዘ ዙ ዚ ዛ ዜ ዝ
ዞ ዟ ዠ ዡ ዢ ዣ ዤ ዥ
ዦ ዧ የ ዩ ዪ ያ ዬ
ይ ዮ ዯ ደ ዱ ዲ ዳ ዴ
ጀ ጁ ጂ ጃ ጄ ጅ ጆ ጇ
ገ ጉ ጊ ጋ ጌ ግ ጎ ጏ ጐ ጑ ጒ ጓ
ጠ ጡ ጢ ጣ ጤ ጥ ጦ ጧ
ጨ ጩ ጪ ጫ ጬ ጭ ጮ ጯ
ጰ ጱ ጲ ጳ ጴ ጵ ጶ ጷ
ጸ ጹ ጺ ጻ ጼ ጽ ጾ ጿ
ፀ ፁ ፂ ፃ ፄ ፅ ፆ
ፇ ፈ ፉ ፊ ፋ ፌ ፍ ፎ
ፏ ፐ ፑ ፒ ፓ ፔ ፕ ፖ
Amharic Numerals
፩ ፪ ፫ ፬ ፭ ፮ ፯ ፰ ፱ ፲
፳ ፴ ፵ ፶ ፷ ፸ ፹ ፺ ፻ ፼
Punctuation Marks
፡ ። ፤ ፣ ! ? ( ) " "
Table II.1 Complete list of Amharic Alphabet
79
III. First Version Amharic Braille
ሀ 01:02:04 ኰ 02:03:04
ሇ 01:04:05 ወ 2:3:4:5
ሏ 1:2:4:5 ዏ 01:06
ሗ 01:02:05 ዖ 01:04:06
ሟ 02:04:05 ዞ 1:3:4:6
ሧ 01:03 ዦ 1:4:5:6
ሯ 01:02:03 ይ 01:05:06
ሷ 1:2:3:6 ጀ 1:2:4;6
ሿ 01:03:04 ገ 1:2:4:5:6
በ 1:3:4:5 ጠ 1:2:5:6
ተ 01:03:05 ጨ 1:2:3:5:6
ቸ 1:3:5:6 ጰ 2:4;5;6
ኀ 1:2:3:4 ጸ 2:4:5:6
ኌ 1:2:3:4:5 ፀ 01:03:06
ኔ 1:2:3:4:5:6 ፇ 1:3:4:5:6
ኜ 1:2:3:5 ፏ 2:3:4:5:6
ኤ 01:02:06
TableIII.1 List of basic Amharic Braille characters in the first version
Vowels 01:04 02:05 03:06 01:05 02:04 02:06 02:06
Variant 1st 2nd 3rd 4th 5th 6th 7th
TableIII.2 List of vowel variant for Amharic Braille characters in the first version
80
IV. Second Version Amharic Braille
1st 6th 1st 6th

ሀ 01:02:05 ህ 1:3:4:6 ኰ 02:03:06 ኵ 5 2:3:6
ሇ 04:05:06 ሌ 01:02:03 ወ 2:4:5:6 ው 02:04:06
ሏ Not apply ዏ Not apply
ሗ 01:03:04 ሜ 02:03 ዖ 1:3:5:6 ዛ 2:3:4:6
ሟ 02:03:04 ሤ 5 ዞ 03:05:06 ዣ 05:06
ሧ 1:2:3:5 ሬ 01:02:05 ዦ 1:3:4:5:6 ያ 1:4:5:6
ሯ Not apply ይ 01:04:05 ዲ 1:4:5:6
ሷ 01:04:06 ሼ 01:05:06 ጀ 02:04:05 ጅ 01:02:06
ሿ 1:2:3:4:5 ቄ 04:06 ገ 1:2:4:5 ግ 2:3:5:6
በ 01:02 ብ 04:05 ጠ 2:3:4:5:6 ጥ 1:2:3:5:6
ተ 2:3:4:5 ት 123456 ጨ 01:04 ጭ 03:06
ቸ 01:06 ች 02:05 ጰ 02:03:05 ጵ 3:4:5:6
ኀ Not apply ፀ 1:2:3:4:6 ፅ 3:4:5:6
ኌ 1:3:4:5 ኑ 1:2:4:6 ጸ Not apply
ኔ 03:04:06 ኙ 02:06 ፇ 01:02:04 ፌ 5 1:2:4
ኜ 3 ኡ 03:04 ፏ
ኤ 01:03 ኩ 03:05
Table IV.1 List of basic Amharic Braille characters in the 2nd version
Vowel None 01:03:06 02:04 1 01:05 none 01:03:05
Table IV.2 List of vowel variants for Amharic Braille characters in the 2nd version
81
V. Third Version Amharic Braille
6th 6th
ህ 01:02:05 ኵ 02:03:06
ሌ 01:02:03 ው 2:4:5:6
Q Not apply ዔ Not apply
ሜ 01:03:04 ዛ 1:3:5:6
ሤ 02:03:04 ዣ 03:05:06
ሬ 1:2:3:5 ያ 1:3:4:5:6
ሴ Not apply ዲ 01:04:05
ሼ 01:04:06 ጅ 02:04:05
ቄ 1:2:3:4:5 ግ 1:2:4:5
ብ 01:02 ጥ 2:3:4:5:6
ት 2:3:4:5 ጭ 01:04
ች 01:06 ጵ 02:03:05
ኅ Not apply ፅ 1:2:3:4:6
ኑ 1:3:4:5 ጽ Not apply
ኙ 03:04:06 ፌ 01:02:04
ኡ *1:2:3:5:6 ý 1:2:3:4
ኩ 01:03
* Symbol indicates Braille code change made on the new version
Table V.1 List of basic Amharic Braille characters in the 3nd version
Vowel 02:06 01:03:06 02:04 1 01:05 Independent 01:03:05

Character
Table V.2 List of vowel variants for Amharic Braille characters in the 3nd version
82
VI. Fourth Version Amharic Braille
Punctuation Braille
mark representation
. 3
: 6 and 3
 2:5
/ 5 and 2
 4 and 1
:- 3:6
' 2
& 2:3
# 2:3:6
$ 3:5:6
' 3:5:6 and 3
? 2:3:6
:: 2:5:6
! 2:3:5
… 3 , 3and 3
() 2:3:5:6
[ 6 and 2:3:5:6
] 2:3:5:6 and 3
* 3:5 and 3:5
_ 4:6 and 4:6
 2:4:6, 2:5 and 2:5
 2:5,2:5 and 1:3:5
 4:5:6 and 1:5
 4:5:6 and 3:5
ኡ኏/ወያሜ 3:4
=//= 6 and 3
X 1:3:4:6
$ 4:5
Table VI.1 List of fourth Version Amharic Braille punctuation marks
83
I. Lookup Table
lookup={
'0':" ",'1':"ሀ",'2':"ሁ",'3':"ሂ",'4':"ሃ",'5':"ሄ",'6':"ህ",'7':"ሆ",
'8':"ሇ",'9':"ለ",'10':"ሉ",'11':"ሊ",'12':"ላ",'13':"ሌ",'14':"ል",
'15':"ሏ",'16':"ሐ",'17':"ሑ",'18':"ሒ",'19':"ሓ",'20':"ሔ",'21':"ሕ",
'22':"ሗ",'23':"መ",'24':"ሙ",'25':"ሚ",'26':"ማ",'27':"ሜ",'28':"ም",
'29':"ሟ",'30':"ሠ",'31':"ሡ",'32':"ሢ",'33':"ሣ",'34':"ሤ",'35':"ሥ",
'36':"ሧ",'37':"ረ",'38':"ሩ",'39':"ሪ",'40':"ራ",'41':"ሬ",'42':"ር",
'43':"ሯ",'44':"ሰ",'45':"ሱ",'46':"ሲ",'47':"ሳ",'48':"ሴ",'49':"ስ",
'50':"ሷ",'51':"ሸ",'52':"ሹ",'53':"ሺ",'54':"ሻ",'55':"ሼ",'56':"ሽ",
'57':"ሿ",'58':"ቀ",'59':"ቁ",'60':"ቂ",'61':"ቃ",'62':"ቄ",'63':"ቅ",
'64':"በ",'65':"ቡ",'66':"ቢ",'67':"ባ",'68':"ቤ",'69':"ብ",'70':"ቦ",
'71':"ተ",'72':"ቱ",'73':"ቲ",'74':"ታ",'75':"ቴ",'76':"ት",'77':"ቶ",
'78':"ቸ",'79':"ቹ",'80':"ቺ",'81':"ቻ",'82':"ቼ",'83':"ች",'84':"ቾ",
'85':"ኀ",'86':"ኁ",'87':"ኂ",'88':"ኃ",'89':"ኄ",'90':"ኅ",'91':"ኆ",
'92':"ኌ",'93':"ኍ",'94':"኎",'95':"኏",'96':"ነ",'97':"ኑ",'98':"ኒ",
'99':"ኔ",'100':"ን",'101':"ኖ",'102':"ኗ",'103':"ኘ",'104':"ኙ",'105':"ኚ",
'106':"ኜ",'107':"ኝ",'108':"ኞ",'109':"ኟ",'110':"አ",'111':"ኡ",'112':"ኢ",
'113':"ኤ",'114':"እ",'115':"ኦ",'116':"ኧ",'117':"ከ",'118':"ኩ",'119':"ኪ",
'120':"ኰ",'121':"኱",'122':"ኲ",'123':"ኳ",'124':"ኴ",'125':"ኵ",'126':"኶",
'127':"ወ",'128':"ዉ",'129':"ዊ",'130':"ዋ",'131':"ዌ",'132':"ው",'133':"ዎ",
'134':"ዏ",'135':"ዐ",'136':"ዑ",'137':"ዒ",'138':"ዓ",'139':"ዔ",'140':"ዕ",
'141':"ዖ",'142':"዗",'143':"ዘ",'144':"ዙ",'145':"ዚ",'146':"ዛ",'147':"ዜ",
'148':"ዞ",'149':"ዟ",'150':"ዠ",'151':"ዡ",'152':"ዢ",'153':"ዣ",'154':"ዤ",
'155':"ዦ",'156':"ዧ",'157':"የ",'158':"ዩ",'159':"ዪ",'160':"ያ",'161':"ዬ",
'162':"ይ",'163':"ዮ",'164':"ዯ",'165':"ደ",'166':"ዱ",'167':"ዲ",'168':"ዳ",
'169':"ጀ",'170':"ጁ",'171':"ጂ",'172':"ጃ",'173':"ጄ",'174':"ጅ",'175':"ጆ", …. Continued on next page
84
'176':"ገ",'177':"ጉ",'178':"ጊ",'179':"ጋ",'180':"ጌ",'181':"ግ",'182':"ጎ",
'183':"ጠ",'184':"ጡ",'185':"ጢ",'186':"ጣ",'187':"ጤ",'188':"ጥ",'189':"ጦ",
'190':"ጨ",'191':"ጩ",'192':"ጪ",'193':"ጫ",'194':"ጬ",'195':"ጭ",'196':"ጮ",
'197':"ጰ",'198':"ጱ",'199':"ጲ",'200':"ጳ",'201':"ጴ",'202':"ጵ",'2403':"ጶ",
'204':"ጸ",'205':"ጹ",'206':"ጺ",'207':"ጻ",'208':"ጼ",'209':"ጽ",'210':"ጾ",
'211':"ፀ",'212':"ፁ",'213':"ፂ",'214':"ፃ",'215':"ፄ",'216':"ፅ",'217':"ፆ",
'218':"ፇ",'219':"ፈ",'220':"ፉ",'221':"ፊ",'222':"ፋ",'223':"ፌ",'224':"ፍ",
'225':"ፏ",'226':"ፐ",'227':"ፑ",'228':"ፒ",'229':"ፓ",'230':"ፔ",'231':"ፕ",
'232':"ቨ",'233':"ቩ",'234':"ቪ",'235':"ቫ",'236':"ቬ",'237':"ቭ",'238':"ቮ",
'239':"ሎ",'240':"ሖ",'241':"ሞ",'242':" ሦ",'243':"ሮ",'244':"ሶ",'245':"ሾ",
'246':"ቈ",'247':"ቧ",'248':"ቷ",'249':"ቿ",'250':"኉",'251':"ና",'252':"ኛ",
'253':"ኣ",'254':"ክ",'255':"ዝ",'256':"ዥ",'257':"ዴ",'258':"ጇ",'259':"጑",
'260':"ጧ",'261':"ጯ",'262':"ጷ",'263':"ጿ",'264':"ፎ",'265':"ፖ",'266':"ቯ",
'282':"@",'283':"፩",'284':"፪",'285':"፫",'286':"፬",'287':"፭",'288':"፮",
'289':"፯",'290':"፰",'291':"፱",'292':"፲",'293':"0",'294':"1",'295':"2",
'296':"3",'297':"4",'298':"5",'299':"6",'300':"7",'301':"8",'302':"9",
'304':"፣",'305':"…",'306':"፤",'307':"᎐",'308':"/",'309':"፦",'310':"–",
'311':"*",'312':"!",'313':"፡",'314':"‹…",'315':"*…+",'316':"—",'317':"?",
'318':"«",'319':"።",'320':"»",'321':"…›",'322':"(",'323':")",'324':"**",
'325':"+",'350':"\n"
85
II. Sample Amharic Braille And Translation Result
Left - Sample double sided Amharic braille
Top right - front page translation of left image
Bottom right - back page translation of left image
86
Declaration
I declare that the thesis is my original work and has not been presented for a degree in any other
university.
_________________
June 5, 2015
This thesis has been submitted for examination with my approval as university advisor.
_________________
Dr. Dereje Teferi

Amanuel Worndimu 2015

Uploaded by

Copyright:

Available Formats

Amanuel Worndimu 2015

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Amanuel Worndimu 2015

Uploaded by

Copyright:

Available Formats

Addis Ababa University

College of Natural Sciences

A Generalized Approach to Amharic

A GENERALIZED APPROACH TO AMHARIC

BY: AMANUEL WENDMU

ADVISOR: DEREJE TEFERI (PHD.)

A Thesis Submitted to the School of Graduate Studies of the Addis

A GENERALIZED APPROACH TO AMHARIC

BY: AMANUEL WENDMU

Name and Signature of the members of examining board

List of Tables ....................................................................................................................................vii

List of Algorithms ............................................................................................................................ viii

List of Abbreviation ...........................................................................................................................ix

CHAPTER ONE ................................................................................................................................... 1

1.1. Background ........................................................................................................................ 1

1.1.1. Braille Writing System ................................................................................................ 1

1.1.2. Braille Recognition Systems ........................................................................................ 3

1.2. Statement of the Problem .................................................................................................. 5

1.3. Objective of the Study ........................................................................................................ 8

1.3.1. General Objective ....................................................................................................... 8

1.3.2. Specific Objectives ...................................................................................................... 8

1.4. Scope and Limitation of the Study ...................................................................................... 8

1.5. Methodology of the Study .................................................................................................. 9

1.5.1. Literature Review ....................................................................................................... 9

1.5.2. Dataset Collection .................................................................................................... 10

1.5.3. Implementation tool................................................................................................. 10

1.5.4. Testing Procedure .................................................................................................... 11

1.5.5. Tools Used ......................................................................................................................... 12

1.6. Significance of the Study .................................................................................................. 13

1.7. Organization of the study ................................................................................................. 14

CHAPTER TWO ................................................................................................................................ 15

2.1. History of Braille............................................................................................................... 15

2.3. Double-Sided Braille ......................................................................................................... 19

2.4. Amharic Writing and Braille System.................................................................................. 20

2.4.1 Amharic Writing System ........................................................................................... 20

2.4.2 Amharic Braille ......................................................................................................... 24

2.5. Automatic Braille Recognition .......................................................................................... 27

2.5.1 Image Acquisition ..................................................................................................... 27

2.5.2 Preprocessing ........................................................................................................... 27

2.5.3 Braille Dot Detection ................................................................................................ 28

2.5.4 Braille Cell Formation ............................................................................................... 29

2.6. Review of Braille Recognition Works ................................................................................ 29

2.6.1 Review of Amharic Braille Recognition works ........................................................... 29

2.6.2 Review of double-sided Braille recognition works ..................................................... 32

CHAPTER THREE .............................................................................................................................. 37

AMHARIC BRAILLE RECOGNITION .................................................................................................... 37

3.1. General Design of Amharic Braille Recognition ...................................................................... 38

3.2. Image Acquisition/Digitization .............................................................................................. 40

3.3. Preprocessing of double-sided Braille Images........................................................................ 40

3.4. Segmentation ....................................................................................................................... 41

3.4.1. Braille Dot Detection ...................................................................................................... 42

3.4.2. Separation of recto and verso ........................................................................................ 42

3.4.3. Braille Cell Grouping ....................................................................................................... 44

3.5. Feature extraction ................................................................................................................ 47

CHAPTER FOUR ............................................................................................................................... 50

4.2. Digitization / Image Acquisition............................................................................................. 51

4.3. Thresholding ......................................................................................................................... 53

4.4. Separation of dots for Double-Sided Braille ........................................................................... 55

4.5. Braille Cell formulation ......................................................................................................... 58