Professional Documents
Culture Documents
Nigus Kefyalew
Nigus Kefyalew
Alphabet Signs
This is to certify that the thesis prepared by Nigus Kefyalew, titled Amharic Sign Language
Recognition based on Amharic Alphabet Signs and Submitted in partial fulfilment of the
requirements for the Degree of Master of Science in Computer Engineering compiles with the
regulations of the University and meets the accepted standards with respect to originality and
quality.
Declared by:
Name: Nigus Kefyalew Tamiru
Signature: __________________________________
Date: ______________________________________
Confirmed by my advisor:
Name: Menore Tekeba
Signature: __________________________________
Date: ______________________________________
Abstract
Sign language is a natural language mostly used by hearing impaired persons to communicate
with each other. At present day, sign language explainers are used to eliminate the language
obstacles between people who are hearing impaired and non-impaired one. However, they are
very limited in number. So, automatic sign language recognition system is better to narrow the
communication gap between hearing impaired and normal people.
This thesis work dealts with development of automatic Amharic sign language translator,
translates Amharic alphabet signs into their corresponding text using digital image processing
and machine learning approach. The input for the system is video frames of Amharic alphabet
signs and the output of the system is Amharic alphabets.
The proposed system has four major components: preprocessing, segmentation, feature
extraction and classification. The preprocessing starts with the cropping and enhancement of
frames. Segmentation was done to segment hand gestures. A total of thirty-four features are
extracted from shape, motion and color of hand gestures to represent both the base and derived
class of Amharic sign characters. Finally, classification models are built using Neural Network
and Multi-Class Support Vector Machine.
The performance of each models, Neural Network (NN) and Support Vector Machine (SVM)
classifiers, are compared on the combination of shape, motion and color feature descriptors using
ten-fold cross validation. The system is trained and tested using a dataset prepared for this
purpose only for all base characters and some derived characters of Amharic. Consequently, the
recognition system is capable of recognizing these Amharic alphabet signs with 57.82% and
74.06% by NN and SVM classifiers respectively. Therefore, the classification performance of
Multi-Class SVM classifier was found to be better than NN classifier.
Key Words: - Amharic Sign Language, Fourier descriptor, Neural Network and Support Vector
machine.
Dedicated to
My Grandmothers
And
My Grandfathers
Acknowledgment
First of all, I would like to thank almighty God and his Mother for giving me the strength, peace
of my mind, and good health to achieve whatever I have achieved so far and for guiding me all
the way through.
I would like to express my sincere gratitude to my advisor Mr. Menore Tekeba for his consistent
follow-up and his willingness to offer me his time and knowledge from the inception to the
completion of this thesis. My sincere thanks also goes to staff of Menelik the II Preparatory
School who have dedicated their time for supplying different materials or equipment when they
were needed. Some special thanks goes to Mrs. Habtam and her some students for capturing the
data set. I also wish to express my gratitude to the IT staff members for providing a training
class.
My thanks extend to Daniel Zemene who has spent many hours for proof, experiment, reading
and discussion. Thank you Hassen Seid for your constructive comments on my report.
The final word of thanks goes to people who are not mentioned in name but whose support
helped me complete the study successfully. Thanks to all.
Table of Contents
CHAPTER ONE: INTRODUCTION .......................................................................... 1
3.1 INTRODUCTION................................................................................................... 24
3.2 SIGN LANGUAGE RECOGNITION SYSTEMS FOR FOREIGN LANGUAGES .................. 24
3.3 AMHARIC SIGN LANGUAGE RECOGNITION SYSTEM ............................................. 28
3.4 SUMMARY .......................................................................................................... 30
i
CHAPTER FOUR: DESIGNING AMHARIC SIGN LANGUAGE RECOGNITION
SYSTEM ...................................................................................................................... 31
4.1 INTRODUCTION................................................................................................... 31
4.2 THE PROPOSED ASLRS ARCHITECTURE.............................................................. 31
4.3 SIGN LANGUAGE VIDEO ACQUISITION ................................................................ 33
4.4 VIDEO TO FRAME CONVERSION .......................................................................... 33
4.5 IMAGE PREPROCESSING ...................................................................................... 34
4.6 SEGMENTATION OF SIGN CHARACTER ................................................................. 35
4.7 FEATURE EXTRACTION ....................................................................................... 36
4.8 TRAINING........................................................................................................... 41
4.9 MODEL CONSTRUCTION...................................................................................... 43
4.10 TESTING ............................................................................................................. 44
4.11 SUMMARY .......................................................................................................... 44
5.1 INTRODUCTION................................................................................................... 46
5.2 DATA SETS ......................................................................................................... 46
5.3 IMPLEMENTATION .............................................................................................. 48
5.4 EVALUATION...................................................................................................... 49
5.5 TEST RESULTS .................................................................................................... 51
5.6 DISCUSSION ....................................................................................................... 58
REFERENCES ............................................................................................................ 63
APPENDIX .................................................................................................................. 68
ii
List of Tables
Table 2.1: The six forms of ‘ሀ’ and its Trajectories with types of motions ...................... 11
Table 4.2: Extracted Shape, Motion and Color features for the sign ‘ ሀ’. ....................... 41
Table 5.2: Test Result for Each Fold, NN and SVM ....................................................... 51
Table 5.3 : Accuracy, Precision, Recall and F-Score measure for NN Model. ............... 53
Table 5.4 : Accuracy, Precision, Recall and F-Score measure for SVM Model. ............. 55
iii
List of Figures
Figure 2.1: Amharic Sign Language Finger Spelling ..................................................... 10
Figure 4.1: Architecture of the Proposed Amharic Sign Language Recognition System . 32
Figure 4.6: Fitting centroids to best adjusted R-square for ‘ሂ’sign Character. .............. 38
Figure 4.7: Fitting centroids to best adjusted R-square for ‘ሃ’sign Character. .............. 39
Figure 4.8: Fitting centroids to best adjusted R-square for ‘ሄ’sign Character ............... 39
Figure 4.9: Fitting centroids to best adjusted R-square for ‘ህ’sign Character ............... 39
Figure 4.10: Neural Network Model with One Hidden Layer ....................................... 42
Figure 5.3: NN and SVM classification accuracy results with bar chart ........................ 52
iv
Figure 5.4: Model Comparison Using Different Performance Metrics .......................... 57
v
List of Algorithms
Algorithm 4.1: An algorithm to prepare the binary frame .............................................. 36
vi
List of Acronyms
AMSL Amharic Sign Language
ANN Artificial Neural Networks
ArSL Arabic Sign Language
ASL American Sign Language
BASL Bangla Sign Language
BSL British Sign Language
CHSL Chinese Sign Language
ENAD Ethiopian National Association for the Deaf
ETHMA Ethiopian Manual Alphabet
ETHSL Ethiopian Sign Language
FD Fourier Descriptor
HCI Human Computer Interaction
ISL Indian Sign Language
KNN K-Nearest Neighbor
MS Manual Sign
NMS Non-Manual Sign
NN Neural Network
RGB Red Green Blue
SASL South Africa Sign Language
SLR Sign Language Recognition
SVM Support Vector Machine
2D Two Dimensional
vii
Chapter One: Introduction
1.1 Background
Language is a communication tool that is useful for different people. People express their ideas
to the community in which they are living. One of the types of language is sign language which
is a manual form of communication language used in the hearing challenged (hearing impaired)
community [1]. When we say hearing challenged community, it comprises hearing challenged
people, their families and other people who communicate with them. Sign language enhances
communication and transfer information among the members of hearing impaired community.
Sign Language is also needed for formal teaching and learning process for the hearing
challenged students. In addition to these issues, those people use sign language to transfer their
culture from one generation to the other.
Usually, sign language is identified by the country where it is used e.g., American Sign
Language (ASL), British Sign Language (BSL), Indian Sign Language (ISL) and Ethiopian Sign
Language (ETHSL) [2]. By its nature, it is not universal because each country has its own sign
language. Even we may obtain differences in sign languages among regions of the same country
since like spoken language, it is unique to a culture and has evolved over time. In other idea,
some people may think that sign language is a signed version of spoken language, but this is
completely a miss assumption because sign language by itself is a language that has its own
finger spelling alphabet, grammar and vocabulary structure [2].
In Ethiopia, there are a lot of hearing impaired people [3]. These people can communicate each
other and the hearing people by using Amharic sign language (AMSL), with reading and writing
techniques [1, 2]. Relative to communication with reading and writing techniques,
communication through AMSL is more preferable for hearing challenged to hearing challenged
people, hearing challenged to hearing people and hearing people to hearing people. However,
there is communication gap among those people because they don’t have the skill of Amharic
sign language. Therefore, to reduce the short coming of the communication based on Amharic
sign language, the research work about AMSL may be very useful since it would make
communication to be good among them.
1
Including hearing impaired people into the common work was very difficult in the world, mainly
due to lack of communication. However, the gap is partially solved by different researchers who
have been researching on methods to develop different applications about different sign language
[4]. However, less developments and research works have been done in Amharic sign language.
Moreover, the field is still hot research area and is not matured well. We believe that working on
this issue will contribute to the growth of the research and development of Amharic sign
language tools.
According to [6], there are 8% disabled people in Ethiopia. Out of these, 2% counts for hearing
impaired people. Irrespective of this number, still there is a clear communication gap between
hearing impaired and hearing people. In order to fill this gap, different measures should be
considered; among those measures sign language training to hearing people and/or coming up
with a system that facilitates the communication of theirs.
Concerning to the second option different research works have been conducted to come up with
a system that converts sign language to text or vice versa for different sign languages all over the
world. For example, American Sign Language (ASL) [7, 8, 9], Indian Sign Language (ISL) [10,
11], South African Sign Language (SASL) [12], Chinese Sign Language (CHSL) [13] and
Bangladesh Sign Language (BASL) [14]. In our country, Legesse Zerubabel [15] has attempted
to develop a recognition system for Amharic alphabet signs which translates a given alphabet
sign into text. The drawback of this work is that it only focuses on the recognition of selected ten
basic alphabet signs from static images; and hence further research work should be done at least
to recognize all the basic and some derived alphabet signs from video of Amharic sign language.
Even though different required resources for sign language recognition in Ethiopia are in their
infant stages, the main aim of our work is to come up with a recognition system that translates
2
the basic and some derived Amharic alphabet signs into equivalent alphabet. So, the system tried
to investigate the fundamental problems that exist between the hearing impaired and hearing
people. To this end, our work discusses the answer to the following research question:
• Can we have a system that correctly recognizes the basic and some derived Amharic
alphabet signs and translate it to the alphabet from?
1.3 Objectives
General Objectives
The general objective of this study is mostly to implement and test Amharic alphabet signs
recognition system using NN and SVM for all the basic and some derived Amharic alphabets
from their corresponding signs.
Specific Objectives
The specific objectives are:
• To prepare the data set or corpus from different signers for this thesis and future
researchers.
• To implement and test Amharic sign alphabet recognition system using NN and
SVM machine learning tools.
• To provide part of a solution to lessen the communication gap between hearing
and hearing challenged people.
• To enhance the computational resources of Amharic sign language recognition
system.
3
such as American Sign Language, Indian sign language, South African sign language,
Bangladesh sign language and Chinese sign language. Based on the information that is obtained
from these papers, the tools and the Algorithms were selected to develop the application.
In addition, to get broader understanding and also to insure the recognition or classification
accuracy, we studied Amharic alphabet signs in detail.
Data Collection
In the course of this study, the first task is to follow how the signers spell Amharic alphabet signs
one by one so that we have an understanding of the Amharic alphabet signs. Hence, this task is
carefully employed in order to address the research objective and come up with the recognition
system. Then we have used a mobile device to record videos of the Amharic alphabet signs
which are used as data set to train as well as test the recognition system.
Tools
Unlike the video capturing part of the proposed system, other parts such as, preprocessing,
segmentation, feature extraction and classification process were done by MATLAB built
function.
Micro soft visio 2010 and Microsoft office 2016 software were used for designing the system
architecture and various tables or diagrams to prepare the document respectively.
Prototype Development and Evaluation
To evaluate the system, the samples data which are collected from different signers were feed
into the developed prototype. Accordingly, the system is evaluated by comparing its output
against the actual classes of Amharic alphabet signs. After the prototype was tested with sample
signs the accuracy of the classification and verification was calculated. The result will be
analyzed and evaluated from which a conclusion and further works will be recommended.
4
This work successfully incorporates the first order and some derived Amharic alphabet signs
translation into their corresponding alphabet texts. The system anticipated to provide various
benefits for both hearing challenged and normal people. Some of the benefits include:
• The key aim of this research work is to come up with a recognition system for ETHSL.
This has an important contribution towards reducing the communication gap that
happens between hearing challenged and hearing persons.
• The system supports the hearing challenged individuals to be participated in community
services for using such kind of applications, as it did for the hearing individuals.
• This current work supports the students to quickly learn Amharic alphabet signs because
sign language training process in the school consumes a long time.
• As an academic exercise, this research increases the experience of other researchers
about ETHSL in the research area.
• The corpus collected and prepared can help future researches in this direction.
5
1.7 Organization of the Rest of the Thesis
The remaining part of the thesis is organized as follows. The basic theory, concepts and the
different issues related to Sign Language (Ethiopian Sign Language and non-Ethiopian Sign
Language) and other relevant topics of its classification and recognition for better understanding
of our research domain are discussed in Chapter two. A general overview characteristic of
Ethiopian Sign Language is discussed in the same chapter. Finally, this chapter presents about
the main parts of Amharic sign language recognition system such as, preprocessing,
segmentation, feature extraction and classification.
Chapter three reviews related research works that has been done on the classification and
recognition of ASL, ISL, CHSL, SASLand BASL by using Machine Learning, Specially SVM
and NN. Lastly, the recognition systems that were done on Ethiopian Sign Language are
reviewed.
Chapter four gives a detailed description of the architecture and design issues of our System. The
basic activities of the system, their main operation and the specific activities are presented in this
chapter.
In Chapter five, the implementation of the proposed system architecture and experimental results
are discussed.
Finally, the conclusion was presented from the research discussion (study) and the
recommendation, and some future works are provided in Chapter six.
6
Chapter Two: Literature Review
In this chapter points related with sign language in general, Amharic sign language in particular,
sign language recognition systems using image processing methods, the tools and techniques
implemented in each phases of the recognition system represented.
In different countries, no sign language is the same. Each country developed its own sign
language [1, 16]. Thus, different countries in the world have their own sign language, such as
United States of America has American Sign Language (ASL), India has Indian Sign Language
(ISL), United Kingdome has British Sign Language (BSL), South Africa has South African Sign
Language (SASL), Austria has Austrian Sign Language (AuSL) and Ethiopia has Ethiopian sign
language (ETHSL). This means that the communication among different countries’ hearing
impaired persons can be as difficult as the communication among different countries’ Hearing
persons.
The communication through spoken language is mainly conducted by verbal communication (the
spoken words and their sounds). However, the communication through sign language proceeds
by gestures. Therefore, sign language can be only understood by considering the gestures alone.
Sign language has two basic components which are Manual Signs (MS) and Non-Manual Signs
(NMS) [17, 18, 19].
7
2.1.2 Non-Manual Signs
Even though manual signs include a large portion of sign language vocabulary, Non-manual
signs (NMS) are significant to convey the information of signs. Unlike manual signs which are
performed by hand and arms only, non- manual signs are conducted by facial expressions and
also by body movements. These include head movement, facial expression (smile and anger),
raised eye borrow, eye shift and so on. Facial expressions can be categorized in to two parts
which are lower and upper facial expressions. The former conveys information about a specific
sign by using the mouth area such as, check and lips. The later used head and body movement
for information expression and focuses on a sign or sentence types (i.e., negation, question, etc.)
[17, 18, 19].
According to the 1994 Housing and Population Census of Ethiopia, there were 1, 90,220 hearing
challenged and hard of hearing people and the most are youngsters [20]. These people live
similarly to any other people within their given cultures; however, they are severing from good
interaction with them. The most hearing challenged youngsters live in countryside areas where
there are no schools for themselves. Because of this the majorities are uneducated and spend
their lives in extreme isolation. They also consider themselves as mentally deficient and evil
because of lack of speech [21, 22]. In towns, more awareness has been made regarding the
hearing challenged people. Even though the resources available are not enough for the number of
students, parents are eager to send their children to schools. This helps the hearing challenged
people to be less isolated.
Initially [1, 6, 23, 15], ETHSL derived from American Sign Language. Even if it was derived
from American Sign Language, there is also some influence of Nordic countries sign language
such as Finish Sign language. Some local signs were created and used in specific hearing
impairment people’ schools in the country and later incorporated into Ethiopian Sign Language.
8
Some of the examples of these words are ‘እንጀራ’, ‘ሀበሻ’,’ሬሳ’, etc. It was first taught in Ethiopia
by American missionary and is based on American Sign Language (ASL) and signed English
[21]. To be suited for Ethiopian culture, it has been modified but may still be intelligible with
ASL. Then, after Ethiopia has developed its own sign language, which comprises finger spelling,
notations, structure and signing hand from hearing challenged people community.
Amharic Sign Langue Finger Spelling
Ethiopian National Association for the Deaf (ENAD) [1, 23, 15] developed Ethiopian Finger
Spelling in 1971 and later gained acceptance by the Ministry of Education. These are called
Ethiopian Manual Alphabets (ETHMAs). According to [1], it has 33 base alphabet signs where
each base alphabet has 6 other variations that are created with the same hand shape or form the
base alphabet followed by unique hand movement for each variation. In 2009, additional one
signed alphabet was set for the Amharic letter Ve/’ቨ’. So, now ETHSL has 34 basic Manual
Alphabets.
Amharic finger spelling is different from American, Britain and Indian Figure spelling. In [24,
25], ASL finger spellings do not require motions for most of the letters to be performed by a
dominated single hand. However, the two alphabet signs (‘J’ and ‘Z’) need movement. In British
[26], finger spelling involves both hands but one letter (‘C’) conducted by a single hand. This is
the opposite to the American Sign Language (ASL) and Ethiopian Sign Language, where letters’
signs are represented by single hand shape. Like ASL, British finger spellings do not need
motion for most of the letters. Indian Sign Language [27] used two handed to show the signs of
lower and upper case English Alphabets and hence it is more complex compared to single
handed ASL. In ISL, the most alphabets do not require a unique hand motion but the rest of two
alphabets (‘H’ and ‘J’) signs need movement.
In Ethiopia, the first orders Amharic alphabets are “Ge’ez” which do not need any movement
and the other orders requires motion to show their own corresponding signs. Like American
Finger Spelling, Ethiopian Finger Spelling uses a single dominated hand for all alphabet signs [2,
23, 15]. In Figure 2.1 [3], the Ethiopian finger spelling with the corresponding Ge’ez alphabet
except ‘ቨ’ is depicted bellow.
9
Figure 2.1: Amharic Sign Language Finger Spelling
As we followed the training class about sign language in Menelik II preparatory school, finger
spelling used to spell words that have no signs. If the word has three characters, it will be
represented by the combination of three spelling signs, one for each letter. For example, for the
word “ደረበ”, the signer firstly would perform the sign for the letter, ‘ደ’, then ‘ረ’ and finally ‘በ’.
In many cases like names of persons, countries, cities and some other common words, their signs
are first letter based. The sign for the name “አበበ” for instance, is the sign for ‘አ’ and touching
part of our face. First alphabet-based signing is not limited to names only however the most
common words like “እንጀራ” is also described by using the sign of ‘አ’and adding some
movement that shows the method of baking “እንጀራ” [1, 28].
Hand shape: The hand fingers particular orientation. When the signer is right handed, he/she
conducts the signs by the right hand. On the other hand, when the signer is left handed, he or she
performs the signs by the left hand. The right and left-handed signers are able to illustrate a sign
on the same letter [28]. Figure 2.2 below depicts the first order Manual Alphabet (ETHMA) of ‘ሀ’
by the right and left dominated hand.
10
a. Right handed sign b. Left handed sign
Figure 2.2: Signs for the letter ‘ሀ’
Hand movement: The hand movement creates variation in the letters to be expressed by a
provided sign. Amharic sign language uses seven orders of alphabet which have similar sign but
differentiated by the movement of the hand shape to make an alphabet [28]. Table 2.1 shows the
movement of hand to create the alphabet variants of ETHMA ‘ሀ”.
Table 2.1: The six forms of ‘ሀ’ and its trajectories with types of motions
Orientation: Orientation is the direction of signs. It is useful on signing because it may change
the meaning. A sign can have one meaning when it moves one way and another meaning when it
moves the other way [28, 29]. For example, to say “give” need the give sign from you to other
body and to say “take” need the same with the reverse direction.
Location: This is the signing area when the sign takes place. The signs which are similar with
other signs only distinguish in their location [29]. For example, signing the word "father" and the
word "mother" in AMSL have the same hand shape except for the sign for "father" is signed
11
locating hand at the fore-head and that of "mother" at the chin. Figure [29] illustrates the Father’s
and Mother’s signs.
Signing Hand
As explained earlier in the above, sign language is a visual language which uses our body parts
to convey meanings. It combines hand shapes, orientation, movement of the hand, and facial
expression. The main component in signing is hand shape which is constructed by changing the
shape of our hands either using both or one hands.
Depending on the signer, the two hands can be classified as dominant and non-dominant hand. If
the person is right handed, his right hand is dominant hand. On the other hand, if the person is
left handed, his left hand is dominant hand [1, 21, 23]. There are three types of signs based on
the use of hands. These are:
• One handed signs: are conducted by a single right or left dominant hand only.
for example, Amharic alphabet signs.
• Symmetric two handed signs: Use both the right and left hands and these hands
move the same way.
• Asymmetric two handed signs: Like symmetric two-handed signs, asymmetric
two-handed signs use two dominant and non-dominant hands however, the
dominant hand moves and the other non-dominant hand remains stationary.
12
Amharic Sign Language Structure
Like Amharic spoken language, Amharic sign language has a common word order in sentence
structure. However, the sentence structure in sign language is different from spoken language.
Most of the time, the general syntax of a sentence in Amharic spoken language is “Subject” +”
Object” +” Verb” but in Amharic sign language is “Time” (optional) +” Topic” +” Comment”
where Time may indicate present, past and future tense. The comment is a word or a phrase that
describe the topic. For example, “እኔ ዳቦ በላሁ”. This sentence is structured based on spoken
language. On the other hand, the equivalent sentence in sign language is “እኔ መብላት ዳቦ”. In this
example, the topic represents “እኔ መብላት” and “ዳቦ” is the comment [21, 23].
13
a. Device based b. Vision based
Figure 2.4: Approaches in SLR
The other main types of SLR are isolated and continuous sign recognition[30]. The former type
is concerned with the recognition of single signs without continuation to the other sign. These
signs can be either static or dynamic. No other sign is done before or after the isolated signs in
continuation. Thus, this type of sign is not affected by the preceding or succeeding sign. In
continuous signing, a complete sentence is recognized containing of various signs conducted one
after the other. The objective is to identify different signs being conducted continuously.
There are varied techniques which used for recognition of sign language. Different authors have
used different techniques according to the nature of sign language and the signs considered. In
this work, Amharic sign language recognition is mainly handles four basic techniques: These are
image preprocessing, Segmentation, feature extraction and classification methods. Let us see
each method in detail below.
2.4 Preprocessing
The first task before proceeding to any image preprocess is the collection of the input data in
appropriate manner. We captured data set in a day that helped us to reduce computational cost
and the noise of data. These are suitable to preprocess the data set easily [15]. After capturing the
video database from different signers, the database should be transferred onto a computer. Then,
the transferred video is automatically converted into the sequence of many frames. To be easy
for further segmentation process, the frames are essentially cropped and enhanced since the
quality of image was poor. Besides, smoothing and sharpening tasks should be performed on the
sequence of frames in this step. Generally, the main aim of pre-processing is an improvement of
14
the image data that successfully removes unwanted distortions or enhances some image features
relevant for further processing task.
2.5 Segmentation
Segmentation refers to the process of partitioning an input image into meaningful segments
(regions/set of pixels) to be easily analyzed for further feature extraction [31]. This means it
successfully separates the foreground object from background. The name of the operation comes
from the result of the operation, which are segments. It is a process found between image pre-
processing and image feature extraction techniques.
Image segmentation is one of the steps involved in the process of recognition in our case
Amharic sign language recognition. For any vision-based application like sign language
recognition, gray scale and binary images hold less information than color images. However,
they need low computational cost, low computational complexity, and low computational time
and also they do not require special hardware device relative to color images [15].
There are some of the well-established segmentation techniques [32]. Here we will discuss about
threshold technique which is one of the important and the simplest technique in image
segmentation. From a grayscale image, thresholding can be needed to form a binary image. This
method replaces each pixel in an image with a black or white pixel if the image intensity Ii,j is
less than some fixed constant T (that is, Ii,j<T), or a white or black pixel if the image intensity is
greater than that constant. The two important types of thresholding are global and local or
adaptive [33]. The global thresholding uses a single threshold value for the entire image based on
an estimation. The local or adaptive thresholding use multiple threshold values for each section
of the image. It has been shown that Niblack’s locally adaptive technique produces the best
result. However, sometimes unnecessary noises such as holes created on the binary image. To
successfully remove those, morphological operations are more appropriate [2]. The most basic
morphological operations are dilation and erosion. Dilation adds pixels to the boundaries of
objects in binary image, while erosion removes pixels on object boundaries.
15
2.6 Feature Extraction
The other most important step in recognition of sign language is feature extraction which is good
to provide feature vectors as input to classifier for sign language recognition. In this step,
important features are extracted after hand segmentation has been carried out successfully.
There are several types of image features that have been used for image classification. Shape,
motion, color etc. are some of the basic image/frame features [32]. Selecting the right set of
features is the crucial key in order to avoid uncertainty in all pattern recognition systems. The
right features should discriminate between any patterns in the sample as an item of one category.
In case of Sign language, it’s necessary to choose features that uniquely identify the shape of
Manual Alphabet Signs [34]. The description of shapes is generally categorized into two groups.
These groups are:
Region Based Shape Description: In a region-based approach [35, 36, 37], all the pixels within a
shape are used to get the shape representation. For image recognition, region-based shape
features extraction approach is better shape representation and be employed to describe non-
connected or disjoint shapes but larger computational complexity than contour-based approach.
Since it combines information across an entire object, rather than exploiting information just at
the boundary points.
Contour Based Shape Description: In contour or boundary-based approach [35, 36], only the
contour information (pixels) of a shape is taken into account to get the shape representation. This
technique is simple, or more popular because of less computational time. The most common
contour-based shape descriptors wavelet and Fourier descriptors. Wavelet descriptors involve
intensive computation due to the dependency of rotation. Fourier descriptors (FDs) improve the
weak discrimination ability of Wavelets descriptors and also easily normalized [35, 36]. They are
derived from the Fourier transform on a shape signature [35, 36, 37, 38]. Shape signature is a
one-dimensional vector describing the outline of a two-dimensional shape. Even there are some
types of shape signatures, the best results were obtained using the complex coordinates.
As the junior signer wants to illustrate signs in real time, problem of slight variation in
angle (rotation), signer hand size (scaling) and translation (direction) need to consider while
extracting the features. These problems are appropriately solved by Fourier descriptors. N points
forming the boundary of a hand shape are required by taking all the pixels occupied by the
16
boundary or taking N samples from the boundary [37, 38]. This can be done by tracing the
boundary counter clockwise. The coordinates of each point in the shape contour can be stated in
the form of (xk, yk) where 0 < k ≤ N-1. The contour can be expressed by coordinate series as
follows:
𝑨(𝒖) = ∑𝑵−𝟏
𝒌=𝟎 𝑺(𝒌)𝒆
−𝒋𝟐𝝅𝒖𝒌/𝑵
(3)
Where u = 0,1,2,3…. N-1 and A(u) is the complex coefficient or Fourier Descriptor (FD) of the
boundary.
The high frequency descriptors contain information about finer details of the shape while the low
frequency descriptors contain information about the general or global features of the shape [38,
39]. However, it is found that all the Fourier coefficients are not necessary to reconstruct the
boundary because the computational process will be complex. Thus, the optimal number of FDs
be selected for hand shape feature extraction.
Motion Features
In addition to hand shape features, different activities were conducted to extract motion
information by following the trajectory of the hands’ centroid. As indicated by Samuel Teshome
[17], a gesture path pattern which holds centroid points (xhand, yhand). So, the trajectory or curve is
determined between two consecutive points from hand gesture path.
17
two types of classification techniques which are basically supervised and unsupervised
classification.
Supervised classification: It is a commonly used technique on the given application and trying to
predict the results for known examples (samples) [41]. It compares its prediction result to the
target answer and learns from its error or mistakes. The aim of this supervised learning is to
generate a classifier model to perform a right prediction for any provided input data.
In order to classify the signs’ features which are given as an input to the system into different
classes, a machine learning classifier is needed [42]. It is a branch of artificial intelligence. The
methods or techniques of a system can be developed for enabling the computer to learn through
machine learning. To perform this work, there are different machine learning classifiers. These
are Hidden Markov Model (HMM), Bayesian Classifier, Template Matching, Boosted Classifier,
K-Nearest Neighbor (KNN), Artificial Neural Network (AAN), Support Vector Machine (SVM)
etc. In this Section, we will only discuss about Artificial Neural Network and Support Vector
Machine pattern classifiers in detail.
18
ANNs are highly distributed interconnections of adaptive non-linear processing elements which
mean they are large set of interconnected neurons. The set of interconnected neurons execute in
parallel to compute the task of learning. Therefore, ANN seems like human brain into two
respects. The former respect is that knowledge is acquired by the network through a learning
process. On the other hand, interneuron connection strengths known as weights are used to store
knowledge, i.e., the weights on the connections encode the knowledge of a network. The neurons
are modeled after the biological neurons and hence they are termed as neural networks(NNs) [32,
44].
Distributed computation of ANN has the advantages of reliability, fault tolerance, high
throughput (division of computation tasks) and cooperative computing. The adaptation is the
ability to alter a system’s parameters according to some rule (normally, minimization of an error
function). Adaptation enables the system to search for optimal performances. The ANN property
of nonlinearity is also important to produce more powerful computation when compared to linear
processing [32]. There are two basic phases in neural network operation. They are training
(learning) phase and testing (recall or retrieval) phase. In the learning phase, data is repeatedly
presented to the network, while weights are updated to obtain a desired response. In testing
phase, the trained network with weights is applied to data that it has never seen.
ANN [43] has three simple sets of rules: multiplication, summation and activation. At the
entrance of artificial neuron, every input value is multiplied with individual weight. In the middle
Section of artificial neuron is sum function that sums all weighted inputs and bias. At the exit of
artificial neuron, the sum of previously weighted inputs and bias is passing through activation
function that is also called transfer function.
Although there exist many representations of ANNs, each one of these networks possesses four
attributes < Nc, W, σ, δ >, where Nc is a finite set of highly interconnected neurons with outputs
n1, n2, . . ., nk; W denotes a finite set of weights which represents the strength w ij of the
interconnection between neurons ni and nj; σ is a propagation rule which shows how the input
signals to a neuron ni propagates through it. A typical propagation rule may be δ (i) = Σnjwij and
δ is an activation function which is usually a nonlinear function like sigmoid function [32]. The
most popular neural network topology is the Multilayer Perceptron (MLP), which is an extension
of the single layer perceptron proposed by Rosenblatt [44]. Multilayer perceptions, in general,
19
are feed forward network, having distinct input, output, and hidden layers. The architecture of
multilayered perceptron with error back propagation network is shown in Figure 2.5 [32].
In an M-class problem where the patterns are N-dimensional, the input layer contains of N
neurons and the output layer contains of M neurons. There can be one or more middle or hidden
layer(s). The above Figure illustrates a single hidden layer case, which is extendable to any
number of hidden layers. The output from each neuron in the input layer is fed to all the neurons
in the hidden layer. No computations are done at the input layer neurons. The hidden layer
neurons sum up the inputs and pass them through the sigmoid non-linearity and fan out multiple
connections to the output layer neurons.
In feed forward activation, neurons of the first hidden layer compute their activation and output
values and pass these on to the next layer as inputs to the neurons in the output layer, which
produce the networks actual response to the input presented to neurons at the input layer. Once
the activation proceeds forward from the input to the output neurons, the network's response is
compared to the desired output corresponding to each set of labeled pattern samples belonging to
20
each specific class, there is a desired output. The actual response of the neurons at the output
layer will deviate from the desired output, which may result in an error at the output layer. The
error at the output layer is used to compute the error at the hidden layer immediately preceding
the output layer and the process continues [32].
In view of the above, the net input to the jth hidden neuron is expressed as:
𝑰𝒉𝒋 = ∑𝑵 𝒉 𝒉
𝒏=𝟏 𝑿𝒏 𝑾𝒊𝒋 + Ѳ𝒋 (4)
𝟏
𝑶𝒋 = 𝒇𝒉𝒋 (𝑰𝒉𝒋 ) = −𝑰𝒉
(5)
𝟏+𝒆 𝒋
Where x1, x2, . . ., xn is the input pattern vector, weights wij denotes the weight between the
hidden layer and the input layer, and Ѳ𝑗ℎ is the bias term associated with each neuron in the
hidden layer. These calculations are known as forward pass. In the output layer, the desired or
target output is set as Tk and the actual output gained from the network is Ok. The error (Tk - Ok)
between the desired signal and the actual output signal is propagated backward during the
backward pass. The equations governing the backward pass are used to correct the weights.
Thus, the network learns the desired mapping function by back propagating the error and hence
the name error backpropagation. The average error E is a function of weight as shown below:
𝟏
𝑬(𝑾𝒋𝒌 ) = 𝟐 ∑𝑴
𝒌=𝟏(𝑻𝒌 − 𝑶𝒌 )
𝟐
(6)
To minimize the error E, we have to find the root of the partial derivatives
Ə𝑬
∑𝑴
𝒌=𝟏 =𝟎 (7)
Ə𝑾𝒋𝒌
Hence, from this we can obtain the value of updated weights as follows
(𝒏𝒆𝒘) (𝒐𝒍𝒅)
𝑾𝒋𝒌 = 𝑾𝒋𝒌 + Ƞ𝜹𝒋 𝑶𝒋 (8)
where η is the learning rate of the hidden layer neurons.
21
sequentially performing a program of instructions, neural networks explore many hypotheses
simultaneously using massive parallelism. Neural networks have the potential for solving
problems in which some inputs and corresponding output values are known, but the relationship
between the inputs and outputs is not well understood or is difficult to translate into a
mathematical function [32, 44].
In a binary classification, given linear separable training set {x1, x2, ..., xn} with their labels {y1,
y2, ..., yn}, yi∊ (- 1, 1), the SVM binary classifier is trained and the optimal hyper plane is
yielded, which separates the data by a maximal margin. The optimal hyperplane divides the data
points into two groups. Points lying on negative side are labeled as -1, and the other points on
positive side are labeled as 1. When a new example is inputted for classification, a label (1 or -1)
is issued by its position with respect to the hyperplane [37, 45].
For the case of data that is inappropriate (non- linearly separable), the SVM first maps the
original data to a higher dimensiona1 space by a kernel function K, such as, Radial Basis
Function (RBF)and polynomial kernels, and then it linearly separates them [37, 45, 46, 47].
Basically, SVM was developed for binary classification and later on it was extended to solve
multi class problems using “one-against-one” or “one-against-all” (one-versus-rest) strategy. The
one-against-one strategy which builds one SVM for each pair of classes. Another popular
method is one-against-all strategy (standard method) that comprises of constructing one SVM
per class, which is trained to distinguish the samples of one class from the samples of all
remaining classes and it is significantly more accurate for classification. In multi class SVM,
multiclass labels are decomposed into several two class labels and it trains the classifiers to solve
22
the problems and the solution of multi class problem is reconstructed from the outputs of
classifiers. Finally, multi-class SVM model generated from known samples [9, 37].
2.9 Summary
In this chapter, general overview regarding to Sign language have been presented. The notations,
finger spelling, signing hand and structure of Amharic sign language are also discussed. In
addition to this, the two basic approaches (device and vision based) of sign language recognition
are introduced. Theoretical backgrounds about digital image processing that are applied to
Amharic sign language recognition system such as preprocessing, segmentation, feature
extraction, and classification are also presented.
Finally, ANN and SVM classifiers which have been used in Amharic sign language recognition
system are described in detail and also their model’s construction is presented.
23
Chapter Three: Related Works
3.1 Introduction
Researchers have been trying to design and develop Sign Language Recognition Systems
(SLRSs) for different languages. This is due to its importance for narrowing the communication
gap that exists between hearing and hearing challenged individuals. In this chapter, related works
on sign language recognition will be presented. The review focused on recognition of American,
Indian, South African, Chinese, Bangladesh and Ethiopian sign languages. Some of the
recognition systems were done on static images while the others were done on videos.
The other researchers attempted to implement a vision based recognition of finger spelled
American Sign Language alphabets in an automated manner [8]. It was designed based on static
gestures which are the selected ten ASL alphabets namely A, B, C, D, G, H, I, L, V and Y for
recognition. The objective of finger spelling recognition system is to describe a simple and
efficient mechanism to translate the sign into text. This proposed system did not use the user to
wear any data gloves for simplification of image processing and it gives human-machine
interaction only through bare hand and also considered white and black background. In the work
of this, feature extraction was the main task. For feature extraction process, Hough Transform
that is an optimal edge detector is engaged since it finds both the locations and the quantity of
features existing in the image and also it is reliable or efficient feature extractor. The other main
task was gesture modeling and classification that was conducted through Support Vector
Machine (SVM). The system is tested on 180 ASL dataset for recognition of individual alphabets
24
and achieved the overall recognition accuracy of 93.88%. Finally, the researchers recommended
that the system to be extended for recognizing the hand gestures from complex background in
real time.
Indian Sign Language Recognition
The work in [10], is aimed to design Indian Sign Language Recognition System for hearing
challenged people to convey their thoughts or ideas. The basic methods for classifying hand
gestures for Human Computer Interaction (HCI) comprises Glove based and Vision based
techniques. The system that uses bar hand gestures for recognition in the vision-based setup
because it is suitable for both single and double handed gestures. As the method is implemented
completely by using digital image processing technique so the user did not wear any special
hardware device to get features of hand shape. Thus, a system recognizes Indian sign language
(ISL) based on hand gestures and it allows the users to interact with the system in natural way.
Then, the system can recognize 36 hand gestures of an Indian sign language which represents the
English alphabets from A to Z and numbers from 0 to 9 via artificial neural network with feed
forward back propagation algorithm which is used to train the network and translate them into
text and voice. Finally, the system has been experimented with different gesture images collected
by a web camera and succeeded a good result with an accuracy of 91.66%.
Additional paper was presented on the performance comparison of three different classifiers for
human computer interaction (HCI) using hand Gestures in ISL [11]. This was developed based
on Zernike moments (ZMs) with three classifiers KNN, ANN and SVM. For feature extraction,
Anand Kulkarni and Sachin Urabinahatti selected Zernike moments which is region based
feature extraction method because ZMs are translation (direction) and size invariant. The
proposed approach is a robust hand gesture recognition system which is presented for
recognizing static alphabet gestures irrespective of the angles in which the alphabet hand
gestures were captured. Thus, the system developed here is real time hand gesture recognition.
The main problem of gesture recognition depends in the complexity of the classification
Algorithms, especially when using high dimensional feature vectors which become necessary in
order to be able to distinguish several of gestures. Comparative study was carried out to show
which classifier works better in reorganization of gestures among KNN, ANN and SVM. The
comparative results of three classifiers are depicted in Table 3.1 [11] below.
25
Table 3.1: System Efficiency Using Three Different Classifiers
KNN 77.5%
ANN 82.5%
SVM 91%
As we see the above Table 3.1, the results show a good significant accuracy in this real time
recognition system and we can conclude that the system works with better efficiency considering
SVM as a classifier either than ANN or KNN classifiers.
South African Sign Language Recognition
Vision-Based Static Hand Gesture Recognition System using SVM were proposed by S. Naidoo,
C.W. Omlin and M. Glaser [12]. The major aim of their research is to develop a system that will
enhance the interaction between hearing challenged and hearing communities in South Africa.
South Africa has two groups of hand gestures, namely static and dynamic. A static gesture is a
specific hand shape and represented by a single image. A dynamic gesture is a moving gesture,
represented by a sequence of images. However, their research work focused on the design of a
system that will recognize static hand images against complex backgrounds based on South
African Sign Language (SASL). SVM classifier was used to classify hand postures as gestures,
due to its high generalization performance without the need to add a priori knowledge, even
when the dimensions of the input space are very high. The experimental results showed that the
system has produced the recognition rate above 90%.
Chinese Sign Language Recognition
In [13], the researchers developed Chinese Sign Language Recognition for a Vision-Based
Multi-Features Classifier System. They focused on the Chinese manual alphabet which is
composed by 30 hand gestures. At first, extracting features of letter images is done, and then
classification method of SVMs for recognition is brought into use. Fourier descriptors and other
multi-features were introduced in this work. The five image descriptors to describe numerous
visual (color histogram and 7Hu moments) and geometrical (48 dimensional Gabor wavelet, 128
Fourier descriptors, scale invariant feature transform (SIFT)) properties of images. The first two
26
descriptors are computed from every pixel of the whole image, while the others descriptors are
computed from small localized interesting image.
The researchers collected 195 images for each letter, 5850 in all by using a camera device.
Experimentation with 30 classes of the Chinese manual alphabet images was conducted and the
results proved that the features, such as Fourier descriptors, are simple, efficient, and effective to
distinguish hand shape, and the SVMs method has excellent classification and generalization
ability in solving learning problem with small training set of samples. This system was able to
recognize images with 95.0256% accuracy when trained with 1500 images and tested with 4350
images. Finally, researchers recommended the idea of using videos as an input to a recognition
system; obtaining more complete feature extractions for sign language image similarity
characterization; and using Multi-kernel SVMs as classifier [13].
Bangladesh Sign Language Recognition
Among the global/international research reviewed, Md. Atiqur Rahman, Dr. Ahsan Ambia, Md.
Ibrahim Abdullah and Sujit Kumar Mondal [14] worked a research work titled Recognition of
Static Hand Gestures of Alphabet in Bangla Sign Language (BASL). The system has the ability
to recognize 36 selected letters of BASL alphabet using ANN with the idea of it being popular in
speech recognition as well as in handwriting recognition. ANN was trained with features of sign
alphabet using feed-forward back-propagation learning algorithm. This recognition system used
the images of the bare hand for the recognition rather than any gloves. There are 23 images of
each sign of BASL used in the recognition. Therefore, a total of 828 signs are collected. Among
828 images, 540 samples were used for training and the remaining 288 images were used for
testing. The average recognition accuracy of the proposed system is 80.902%. Using more
images or samples for training ANN may improve the performance of the system. The limitation
of this paper is that, the feature vectors should have integer values only. Future work would
include extending the developed method to recognition of BASL with video based system [14].
Md Azher Uddin and Shayhan Ameen Chowdhury [49] proposed another framework for
recognizing Bangla Sign Language (BASL) using SVM. Here, first the original RGB images are
transformed to Hue, Saturation and Value (HSV) color space. Then, features are extracted from
the segmented image by Gabor filter and then Kernel Principal Component Analysis (KPCA) is
applied to reduce the dimensionality since KPCA is a nonlinear dimensionality reduction
technique. Finally, SVM is used to identify alphabets of the sign. In the experiments, they have
27
used 4800 images with the size 480×360 pixels. The images were captured from different
illuminations. From this database, 2400 images were used for training and 2400 images were
used for testing. At the end, the system attained 97.7% accuracy result.
The other work was conducted by Tefera Gimbi [23] in 2014. Unlike the above work [6] which
is rule based system, his work is a machine learning based system that is Isolated Signs
Recognition System in ETHSL. The system receives videos of Amharic word signs and produce
extracted frames. On the sequence frames, skin color detection algorithm was implemented and
the equivalent binary image is created that has white value on the foreground for skin color and
black value on back ground for another region. Based on the detected skin regions the hands and
head were segmented from other parts of the body since they have very important role in the
signing process of Amharic words. Significant features which help to get more shape information
and get a better recognition result are extracted from the segmented body parts. Then, Hidden
Markov Model (HMM) was trained by using these feature vectors. Three signers participated to
capture the data set in this work and each signer performed a sign twenty times. Out of these
twenty videos, Tefera Gimbi used fifteen for training and the rest five video for testing purpose.
The system has been tested using the videos which were collected for learning purpose. He
attained the overall recognition of 86.9% using the HMMs learned by eight features (i.e., Area,
Centroid, Bounding Box, Major Axis, Minor Axis, Eccentricity, Orientation and Perimeter)
28
whereas by using only three basic features (i.e., Centroid, Area and Orientation), he got the
recognition of 83.5%.
According to [20], the two authors attempted to design a hand gesture recognition system by
using Gabor Filter (GF) together with Principal Component Analysis (PCA) for feature
extraction and ANN for recognizing the ETHSL of 34 letters of ETHMA. For the proposed
system, 170 images were captured from five ETHSL students for 34 Amharic alphabets. In order
to reduce the difficulty of segmentation due to high variation of skin color among the signers, the
system used white glove while capturing the total dataset. Moreover, their work describes that
non- ETHEMA images were captured from websites and the system is capable of rejecting these
non ETHMA images. Generally, the basic purpose of this work is translating those ETHSL
alphabet signs to voice. The experiments performed and proved that with sufficient data, ANN
approach produced good results and able to recognize an unknown input sign (non ETHMA)
very fast. The experimental results show that the system has produced the recognition rate of
98.53% [20]. However, the paper could not provide full translation functionality because it could
not be used to translate basic and specific Amharic alphabet signs to the corresponding
alphabets.
In addition to the above research done by [20], Legesse Zerubabel [15] continued to propose
Ethiopian Finger Spelling Classification System. The system architecture includes different
components which are image preprocessing, feature extraction, segmentation and classification.
The system worked by accepting Amharic alphabet signs as inputs and returning the
corresponding Amharic alphabet text as an output.
Different experiments were performed to get the appropriate feature extractors and pattern
classifier for hand detection and sign classification tasks. Besides, through experiments, the
capability of principal component analysis and harr-like feature with neural network were tested.
Basically, the work focuses on hand detection and sign classification that was used as input for
Ethiopian fingerspelling recognition. According to experiment results, the overall recognition
rate of 88.08% and 96.22% were obtained using neural network with PCA driven feature and
neural network with harr-like feature respectively [15]. This developed system is only limited to
the recognition of ten basic Amharic alphabet signs based on images. Therefore, it is necessary
29
to extend the system so that it handles specific and the remaining basic Amharic alphabet signs
with video data set.
Later, this research work was extended by Abadi Tsegay [2]. He attempted to develop offline
candidate hand gesture selection and trajectory determination system for continuous ETHSL.
This recognition system used extracted candidate ETHMA frames from the video sequence and
that can also determine hand movement trajectories. The system has two basic components
namely:
• Candidate Gesture Selection (CGS) and
• Hand Movement Trajectory Determination (HMTD).
The CGS combines two metrics speed profile of continuous gestures and Modified Housdorff
Distance (MHD) measure and obtained an accuracy of 80.72% for this module. The HMTD is
done by considering each hand gesture centroid from frame to frame and using angle, x- and y-
directions, and it returned a result with an accuracy of 88.31%. In addition to this the system as a
whole has a performance of 71.88%. However, he did not focus on the first and the seventh
orders of Amharic alphabet signs except the second, the third, the fourth, the fifth and the sixth
orders. He hasn’t also worked on recognition of these signs to text.
3.4 Summary
Most of the research works were done in the recognition of Amharic sign language mainly
focused on the translation of sign to its equivalent text but the systems couldn’t be tested directly
to all basic and/or some specific alphabets.
Besides, researchers did not use Support Vector Machine classifier for Amharic sign language
recognition and also did not evaluate the performance between ANN and SVM based on this
area. According to the above research reviews, in other countries like America, India, etc.
different researchers used SVM and ANN for Sign Language recognition and also compare SVM
performance with ANN and achieved good performance. Because of this, we selected ANN and
SVM to be used as classifiers for our research work.
30
Chapter Four: Designing Amharic Sign Language Recognition
System
4.1 Introduction
As described in different research works, recognition of Amharic sign language into the
corresponding character using digital image processing and learning machines passes through
different phases: sign language video acquisition, video to frame conversion, frame
preprocessing, segmentation, and training and recognition. In this chapter, we propose a design
for recognition of Amharic Sign Language into Amharic character. The Sections in this chapter
present details of each component in the architecture. Section 4.2 presents the general overview
of the proposed system architecture. Section 4.3 presents Sign Language video acquisition
techniques and tools used. Section 4.4 explains techniques used to convert video into frame.
Section 4.5 briefly explains frame preprocessing. Section 4.6 explains techniques and algorithms
used in segmenting sign character from the background. 4.7 explains the tool and techniques used
to extract features from the segmented character. After features are extracted, selected learning
machines are trained from which models are constructed and tested. Finally, a summary for the
chapter is briefly presented in the last Section.
The first component converts acquired sign language video to frame images. The frame images
are preprocessed to remove unnecessary noise in the second component. The third component
segments sign character from the background. Fourier descriptor, motion descriptors and color
difference measure are used to extract features in the fourth component. In the training
component, extracted features are feed into learning machines (NN and SVM) to construct their
respective models. The Models constructed are used to recognize sign language into print text in
the testing component. In this component the performance of the models constructed (for NN and
SVM) are also tested.
31
Figure 4.1: Architecture of the Proposed Amharic Sign Language Recognition System
32
4.3 Sign Language Video Acquisition
This is the first step in recognizing Amharic sign language. Before proceeding to any recogniton
processes, all the required videos are collected from which frames are extracted.
33
research work, we set the minimum frame to be not less than 50 frames (i.e play time greater
than 3 seconds). This helps us to have enough number of centroids for each frames collected.
The equation of motion of the frames: linear (left, right, down), sinosidal and nearly circle is
formulated from the position of the centroids of other frames in the video relative to the centroid
of the first frame(reference frame). The positions (x,y) of the centroids of each frames are also
authoamticaly detected by using the region property of the segmented frame image, from which
the motion descripters are drived as broadely described in Section 4.7. Figure 4.3 shows the
number of frames (58 in number) detected from the video for a play time of 3.56 seconds for
character ‘ሆ’.
34
Cropping: The frames which are extracted from the orginal video have huge size and unnesasary
frame components. As a result, processing the sequence of video frames with huge dimension is
computationally heavy for further processing. Hence, the height and the width of orginal frames
are cropped.
Converting RGB image into grayscale form: Firstly, the orginal extracted frame is RGB (24 bits
color) form in which each pixel is presented by Red, Blue, and Green components. On the other
hand, Grayscale image (8 bits image) only has shades of black and white. It doesn't have any
color information. Processing the image in RGB form is computationally heavy and it takes
more processing time when compared with grayscale form. Because of this, the sequenc of video
frames is converted into grayscale representation form for suitable processing.
Contrast adjustment: After the original video frame in RGB color space is converted into
Greyscale color frame, identifing the foreground and the background object by our eye is
difficult. Inhancing the quality and adjusting the contrast of the video frame reduce the effect of
this problem on the performance of the system. Basically, the idea behind this techniques is to
bring out detail that is obscured, or simply to highlight certain features of interest in a frame. A
familiar example of enhancement is when we increase the contrast of an image because it looks
better.
Sharpening: The Sharpening process increases the contrast between bright and dark hand
regions to bring out good features. In applications where edges are very important, an image
sharpening technique is used to enhance edges from the blurred image. In addition to this, it
also differentiate the blurred open or close hand fingers. To increase the discrimination ability
of the neural network and support vector machine model, the input frames should somewhat be
sharp.
We select adaptive threshold algorithm because our data sets are affected by shadows, shading
and lighting effects. Besides, we applied different morphological operators such as dilation,
35
erosion and morphological filling operations to remove tiny objects and re fill missed objects of
the segmented sign.
36
sign characters. These shape features are taken after the original shape is resampled as shown in
Figure 4.5 and they are extracted using Algorithm 3 illustrated below. As described in the
Algorithm, segmented binary frame image is taken as an input, as discussed in Section 2.6 and
the segmented image is resampled to reduce the computational time of the Fourier descriptor. On
the resampled image, we applied a 1D complex signal on which Discrete Fourier Transform
(DFT) is implemented which returns list of feature descriptors (FDs). These list of feature
descriptors are normalized to 31 best feature descriptors.
37
Motion Feature Descriptors
In addition to the shape features extracted from the captured videos, features conveyed by the
hand motions were also considered for this research. As stated in Section 2.6, motion can be
represented in various techniques. In the proposed system, the two motion features are found to
be crucial to represent sign gestures namely, Direction and Angle to identify the type of
trajectory or curve.
Direction describes the type of movement with reference to the four-coordinate plane. Hence,
the direction of sign characters could be identified based on the centroid information, gathered
from a sampled frame lists, followed by analyzing the direction of the last frame with respect to
the first video frame. By this, it is meant that the last frame can be referred as either to the right,
left, or down side of the first video frame.
The type of curve or trajectory might depend on the curve trajectory formed by drawing a line
between each individual centroid and realized by employing different models. For example: A
polynomial fitting with order one or two produce a good result to model for the first three
derivatives (ካዕብ, ሣልስ and ራብዕ). Figure 4.6 and Figure 4.7 shows fitting the centroids, extracted
from ten sampled frames for Amharic sign characters (‘ሂ’) and (‘ሃ’) respectively.
The other derivatives (the Hamis(ኃምስ)-ሄ and the Sadis-ህ) are modeled by circle and sinusoidal
models respectively as shown in Figure 4.8 and 4.9.
Figure 4.6: Fitting centroids to best adjusted R-square for ‘ሂ’sign Character
38
Figure 4.7: Fitting centroids to best adjusted R-square for ‘ሃ’sign Character.
Figure 4.8: Fitting centroids to best adjusted R-square for ‘ሄ’sign Character
Figure 4.9: Fitting centroids to best adjusted R-square for ‘ህ’sign Character
Accordingly, the five derived Amharic sign characters could be model by the following
equations as shown in Table 4.1. From these models, we depict two features to represent the
characters: the slope named as angle difference and direction. Angle refers the angle make by
consecutive centroids of the frames with respect to the centroid of first frame. For the first three
derived Amharic sign characters (ሁ, ሂ and ሃ) the angle difference measure is measured to be less
than 30o in our experiment. However, for the ሄ sign of the angle difference measure is found to
be more than 180o and the rest ህ’s sign, the angle difference is less than 180o in our experiment with
respect to a line drown from the first frame to the last frame in the direction of motion.
39
Table 4.1: Model Equations for Amharic Sign Character
Amharic Sign for Sabi (ሳብዕ) Character = Shape of the first frame + Skin color of the last frame
Input: -colors of first frame and last frame around the centroids
Output: - color difference measure value, CDM
Get collection of RGB colors []c1 from the first frame around its centroid.
Get collection of RGB colors []c2 from the last frame around its centroid
Convert both color collections from RGB to HSV collections.
Calculate average value of each collection: C1avg and C2avg
1
Compute a reference color value from C1avg and C2avg: Cref= 2(C1avg + C2avg)
Compute rms value of C1avg and C2avg relative to Cref: Cd1 and Cd2
i.e. Cd1 =√ (C1avg - Cref)2and Cd2 = √ (C2avg - Cref)2
Compute color difference measure: CDM = |Cd1 – Cd2|
Return CDM
40
Algorithm 4.3 : Color Difference Measure
Combining the above features, 31 from shape feature, 2 from motion feature and 1 from color
feature, we come up 34 features to represent Amharic Sign Language Characters as shown in
Table 4.3. These features are stored in the Knowledge Base (KB) for training.
Table 4.2: Extracted Shape, Motion and Color features for the sign ‘ሀ’.
4.8 Training
In this component extracted features obtained from the feature extraction is used as an input in
performing classification through training. We have selected two classifiers: artificial neural
network (ANN) and Support Vector Machine (SVM) as training machines to construct models
which would be used as classifiers for Amharic sign characters into the corresponding Amharic
print characters.
In the case of ANN, we need to select optimum architecture (topology) and training an algorithm
that best matches to the proposed classification model. In our study, feed forward multi-layer
41
perceptron (MLPs) architecture is used. It is composed of three types of layers: input layer,
output layer and hidden layer.
To compute the neuron final output state, a backpropagation algorithm needs continuous and
differentiable activation function. For this study, in order to control the smooth relation between
the input and output, a sigmoid transfer function is used.
Generally, as depicted in Figure 4.10, we used feed forward multilayer perceptron with one
hidden layer which has 75 neurons in addition to the input and output layers. The features: shape
feature descriptors, motion feature descriptors and color feature descriptors computed are used as
an input layer in NN architecture. Based on these selected features, the number of input neurons
is 34 for the selected 52 Amharic alphabet signs, as a result NN gives us 52 generated output
classes in the output layer as shown in Figure 4.10.
To build a model using SVM which is basically binary classifier, we applied a multi class SVM
classifier since the expected outputs generated from our model has to be more than two classes as
shown in Figure 4.11. Though, there are various multi-class SVM with kernel functions which
42
are used to construct a model for classifying Amharic sign characters, one against all with radial
basis function is found to be better classifier to identify actual multi-classes for Amharic sign
language with large number of signs. We used different parameters and initial values to generate
an optimal result.
For NN, as we already illustrate in Figure 4.12, the optimal number of hidden layer neurons and
the size of the layers in the architecture are determined by doing several training steps iteratively.
As a result, we finally constructed a NN model with 34 different input neurons, 75 neurons on
the hidden layer to determine one of the 52 classes. Figure 4.12 shows the final generated model
of NN.
In the case of SVM, the training steps are selecting different parameters along with its
corresponding initial values iteratively. We selected “One against All” method which is an
optimal to construct 52 classes by taking 34 features as an input as can be seen in Figure 4.11.
43
Figure 4.12: Constructed Model for NN
4.10 Testing
In this component, the recognition or classification accuracy of both constructed models: NN and
SVM are tested. As early presented in Section 2.8, k-fold cross validation is an appropriate
method to test NN and SVM models. Hence, we select k-fold cross validation technique in this
work because all data set are used for both training and validation. In our case the value of k is
ten. The whole data set is partitioned into 10 equal parts and each partition is used for both
training and evaluation. As described in Table 5.1, the total data that are 1,710 from which
58,140 set feature vectors are feed into the training machines.
4.11 Summary
This chapter systematically went through the designing of Amharic sign language recognition
system.
The proposed system has different components for systematically working together. Initially, for
each Amharic sign language characters, sample videos are acquired. Collection of frames are
taken from the videos. The frames are preprocessed and segmented from the background. Then,
after preprocessing and segmentation, from the first frame of each character sign, we take shape
features using Fourier Descriptor. From the other consecutive set of frames, motion and color
features are derived in the feature extraction components. From this component, thirty-four
44
different features are extracted. Among these features 31 of them are shape features extracted
from the first frame. Others are motion and color features extracted from the motion of other
frames and the color of the last frame respectively. Finally, an artificial neural network and
support vector machine are trained using these features from which 52 output classes are
generated to classify or recognize the selected Amharic Sign language characters. The sign
characters are the 34 basic Amharic sign language characters. And the others are derived
Amharic sign characters of ‘ሀ’,’ለ’ and ‘ሐ’. Through training, in this work two models (NN and
SVM) are constructed and then tested. The models are tested using ten-fold cross validation
testing technique.
45
Chapter Five: Experimentation and Result Discussion
5.1 Introduction
In the previous chapter, we described the design of the system. This chapter discusses
implementation detail of the proposed design for recognition of Amharic sign language. Section
5.2 presents the datasets used in training and testing the system. Section 5.3 describes the tools
used and the overview of the system. Section 5.4 presents the evaluation and test results found in
Section 5.5. Finally, the discussion part is presented in the last Section of the chapter.
As depicted in the Table 5.1, total sample videos were collected for basic Amharic alphabet signs
which are 1,530 and also total samples are 180 were captured for some derived Amharic alphabet
signs. Accordingly, we have created our own data set that contain around 1,710 real videos of 52
different classes that are basic and some specific manual alphabets.
46
Table 5.1: The collected ETHSL manual alphabets
47
In the process of data collection, a lot of challenges have been occurred to address about
Amharic alphabet signs with different signers (teachers) in Menelik the II Preparatory School.
Unfortunately, the reason that school’s signers are not willing for giving information was the
main challenges. The challenges have been made their own impact on the success of our work,
especially, during recording. Besides, getting notes and books about ETHSL was also a very
challenging task.
5.3 Implementation
We used MATHLAB 2014a and MySQL to implement the prototype. MATHLAB is a high-
level language and interactive environment used to design and test the tools and techniques used
in our approach. MySQL is used to persist the features extracted from Amharic sign language
characters. Figure 5.1 shows the running prototype. The designed prototype is tested on Toshiba
laptop core i3, processor speed of 1.6 GHz, 4.0GB RAM,700 GB hard disk capacities and 64 bits
operating system.
48
5.4 Evaluation
In many machine learning areas, a basic problem is obtaining an accurate estimate for the
generalization ability of a learning algorithm trained on a given dataset. Hence, the basic concern
in machine learning is to obtain an accurate estimate of the generalization error of a model
trained on a finite dataset [50]. Simply splitting the corpus or dataset into a single training and
testing set may not give the best estimate of future performance. Therefore, we used different
evaluation techniques which is cross- validation with confusion matrix on our approach to
clearly point out the accuracy of the model. This can also help to choose an algorithm from a
variety of learning algorithms.
Most machine learning approaches are evaluated by using a methodology of cross validation
which is believed to be a more reliable methodology [51]. It is a statistical method of evaluating
and comparing learning algorithms or models by dividing data into two segments: one used to
learn or train a model and the other used to validate the model. The basic form of cross-
validation is k-fold cross validation which is the most appropriate technique. “In k- fold cross
validation, the dataset D is randomly split into k subsets D1,D2,… Dk of approximately equal
size” [51].
K- Fold cross validation helps us to estimate performance of the learned models (NN and SVM)
from our dataset to gauge the generalization of the algorithms for our Amharic sign language
classification as it was described in Section 4.10. Thus, we can compare the performance of
models constructed from NN and SVM and find out which one is the best. Generally, it generates
an approximate measurement of how well the learned model will do on “unseen” data.
Our experiment is done using a 10-fold cross-validation on the available data. This means that
the data is split in ten equal partitions, and each of these is used once as test set, with the other
nine as corresponding train set. This way, all examples are used at least once as a test item, while
keeping training and test data carefully separated, and the NN and SVM based classifiers are
trained each time on 90% of the available training data. The process of our 10-fold evaluation is
illustrated in Figure 5.2.
49
Figure 5.2: Process of 10-fold cross validation experiment
The lighter section of the data (1,539 Amharic sign characters’ data set) are used for training
while the darker sections (171 sign characters) are used for validation in each experiment.
In our work, we have used four performance matrices to analyze the classification performance
of NN and SVM models using 10-fold cross validation. These performance matrices are:
Accuracy, Precision (positive predictive value), Recall (True positive rate or Sensitivity) and F-
score. Accuracy is the total number of correctly predicted signs to all test samples. Precision is
the ratio of correctly identified instances of a class to the total number of positive observations
identified a sign as that class. Recall is the ratio of correctly identified instances of a class to all
the instances where the sign was truly that class and f-score is the harmonic mean of precision
and recall [4, 8]. All of these parameters are given by the following equations (9), (10), (11) and
(12) respectively:
𝑻𝑷+𝑻𝑵
𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 = 𝑻𝑷+𝑭𝑵+𝑭𝑷+𝑻𝑵 (9)
𝑻𝑷
𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏 = 𝑻𝑷+𝑭𝑷 (10)
𝑻𝑷
𝑹𝒆𝒄𝒂𝒍𝒍 = (11)
𝑻𝑷+𝑭𝑵
50
𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏∗𝑹𝒆𝒄𝒂𝒍𝒍
𝑭𝑺𝒄𝒐𝒓𝒆 = 𝟐 ∗ (𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏+𝑹𝒆𝒄𝒂𝒍𝒍) (12)
In these equations, TP, TN, FN, FP present true positive, true negative, false negative, false
positive respectively. A true positive occurs when the predicted class is the same as the actual
class. A false positive occurs when a classifier classifies sign character as the incorrect class. A
true negative is when the classifier correctly predicts sign character not part of an incorrect class
and a false negative occurs when the classifier does not classify sign character into the correct
class [4].
Experiment 1 2 3 4 5 6 7 8 9 10 Avg
(%)
Total No of 1710 1710 1710 1710 1710 1710 1710 1710 1710 1710 1710
Dataset
Accuracy- 61.73 52.63 60.17 61.66 59.77 47.81 51.97 59.57 60.29 62.64 57.82
NN (%)
Accuracy- 74.06
73.5 74.93 72.95 75.42 72.71 74.26 72.74 73.6 74.51 75.98
SVM (%)
As can be seen in the Table 5.2, we found that the recognition accuracy of Neural network is
61.73%, 52.63%, 60.17%, 61.66%, 59.77%, 47.81%, 51.97%, 59.57%, 60.29% and 62.64% on
experiment one, experiment two, experiment three, experiment four, experiment five, experiment
six, experiment seven, experiment eight, experiment nine and experiment ten respectively. The
lowest recognition percentage (47.81%) is recorded on experiment six which indicates different
51
signs in this experiment are very similar in shape but the highest recognition percentage
(62.64%) is recorded on experiment ten that means it includes signs in dissimilar shape. The
average recognition accuracy of NN from those experiments is 57.82%.
The other recognition accuracy of Support Vector Machine is 73.50%, 74.93%, 72.95%, 75.42%,
72.71%, 74.26%, 72.74%, 73.60%, 74.51% and 75.87% on experiment one, experiment two,
experiment three, experiment four, experiment five, experiment six, experiment seven,
experiment eight, experiment nine and experiment ten respectively. In experiment five (fold 5),
there is the lowest recognition percentage (72.71%). However, the highest recognition
percentage (75.98%) is recorded on experiment one. The average recognition accuracy from
those ten experiments is 74.06%. In each experiment, the accuracy result of SVM is greater than
NN’s accuracy result. The classification performance of NN and SVM Models in the above
Table 5.2 are clearly illustrated with a Bar chart below.
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10
NN SVM
Figure 5.3: NN and SVM classification accuracy results with bar chart
The Bar chart in Figure 5.3 is a visual representation of the results above. The blue color
represents the Neural Network’s accuracy and the red color represents the Support Vector
Machine’s accuracy. This is a good chart on how well the performance of the NN and SVM is
compared in each fold.
52
Each class result of accuracy, precision, recall and F-score based on NN and SVM models from
the confusion matrix is presented in the Table 5.3 and Table 5.4 respectively as tabulated below.
Table 5.3 : Accuracy, Precision, Recall and F-Score measure for NN Model.
53
ገ 26 31 39 28 74 61.04651163 44.28571429 52.542373 48.0620155
ጠ 27 38 25 28 81 69.18604651 60.31746032 57.575758 58.91472868
ጨ 28 36 33 34 69 61.04651163 52.17391304 51.428571 51.79856115
ጰ 29 32 32 33 75 62.20930233 50 49.230769 49.6124031
ጸ 30 39 29 25 79 68.60465116 57.35294118 60.9375 59.09090909
ፀ 31 34 38 25 75 63.37209302 47.22222222 57.627119 51.90839695
ፈ 32 38 35 24 75 65.69767442 52.05479452 61.290323 56.2962963
ፐ 33 33 34 29 76 63.37209302 49.25373134 53.225806 51.1627907
ቨ 34 37 38 33 64 58.72093023 49.33333333 52.857143 51.03448276
ሁ 35 15 39 35 83 56.97674419 27.77777778 30 28.84615385
ሂ 36 12 34 36 90 59.30232558 26.08695652 25 25.53191489
ሃ 37 11 36 38 87 56.97674419 23.40425532 22.44898 22.91666667
ሄ 38 12 35 39 86 56.97674419 25.53191489 23.529412 24.48979592
ህ 39 11 35 36 90 58.72093023 23.91304348 23.404255 23.65591398
ሆ 40 11 33 38 90 58.72093023 25 22.44898 23.65591398
ሉ 41 10 39 41 82 53.48837209 20.40816327 19.607843 20
ሊ 42 12 44 43 73 49.41860465 21.42857143 21.818182 21.62162162
ላ 43 15 56 54 47 36.04651163 21.12676056 21.73913 21.42857143
ሌ 44 17 43 42 70 50.58139535 28.33333333 28.813559 28.57142857
ል 45 12 56 47 57 40.11627907 17.64705882 20.338983 18.8976378
ሎ 46 13 58 59 42 31.97674419 18.30985915 18.055556 18.18181818
ሑ 47 18 53 55 46 37.20930233 25.35211268 24.657534 25
ሒ 48 14 41 42 75 51.74418605 25.45454545 25 25.22522523
ሓ 49 9 39 34 90 57.55813953 18.75 20.930233 19.78021978
ሔ 50 11 31 34 96 62.20930233 26.19047619 24.444444 25.28735632
ሕ 51 13 31 34 94 62.20930233 29.54545455 27.659574 28.57142857
ሖ 52 11 32 28 101 65.11627907 25.58139535 28.205128 26.82926829
Total 3006.976744 2124.926236 2124.0268 2121.604091
Average (%) 57.82647585 40.86396608 40.846669 40.80007866
54
Table 5.4 : Accuracy, Precision, Recall and F-Score measure for SVM Model.
55
ጨ 28 36 20 24 92 74.41860465 64.28571429 60 62.06896552
ጰ 29 32 23 23 94 73.25581395 58.18181818 58.181818 58.18181818
ጸ 30 29 20 19 104 77.3255814 59.18367347 60.416667 59.79381443
ፀ 31 28 33 21 90 68.60465116 45.90163934 57.142857 50.90909091
ፈ 32 39 31 15 87 73.25581395 55.71428571 72.222222 62.90322581
ፐ 33 31 18 13 110 81.97674419 63.26530612 70.454545 66.66666667
ቨ 34 27 18 14 113 81.39534884 60 65.853659 62.79069767
ሁ 35 12 29 31 100 65.11627907 29.26829268 27.906977 28.57142857
ሂ 36 13 24 24 111 72.09302326 35.13513514 35.135135 35.13513514
ሃ 37 9 26 24 113 70.93023256 25.71428571 27.272727 26.47058824
ሄ 38 12 25 19 116 74.41860465 32.43243243 38.709677 35.29411765
ህ 39 9 20 20 123 76.74418605 31.03448276 31.034483 31.03448276
ሆ 40 8 27 18 119 73.8372093 22.85714286 30.769231 26.2295082
ሉ 41 12 29 18 113 72.6744186 29.26829268 40 33.8028169
ሊ 42 11 34 22 105 67.44186047 24.44444444 33.333333 28.20512821
ላ 43 11 47 24 90 58.72093023 18.96551724 31.428571 23.65591398
ሌ 44 13 31 34 94 62.20930233 29.54545455 27.659574 28.57142857
ል 45 9 22 28 113 70.93023256 29.03225806 24.324324 26.47058824
ሎ 46 12 25 29 106 68.60465116 32.43243243 29.268293 30.76923077
ሑ 47 11 23 38 100 64.53488372 32.35294118 22.44898 26.5060241
ሒ 48 9 21 36 106 66.86046512 30 20 24
ሓ 49 11 28 37 96 62.20930233 28.20512821 22.916667 25.28735632
ሔ 50 12 26 34 100 65.11627907 31.57894737 26.086957 28.57142857
ሕ 51 11 21 34 106 68.02325581 34.375 24.444444 28.57142857
ሖ 52 9 32 20 111 69.76744186 21.95121951 31.034483 25.71428571
Total 3851.162791 2638.429929 2644.5932 2629.803292
Average (%) 74.0608229 50.7390371 50.857561 50.57314023
As indicated in Table 5.3, the summary results of neural network model showed that 57.82%
were correctly classified and 42.18% were misclassified and we can see that the best cases
(above 65%) are the letters ‘መ’, ‘ዐ’,’ጠ’, and ‘ጸ’, mainly because these signs have very few
56
similarities with the others. Conversely, the worst case is the letters ‘ላ’,’ ሎ’ and ‘ሑ’ with only
below 40% accuracy. In other way, it is observed that the precision, recall and f_score results of
NN model are 40.86%, 40.84% and 40.80% respectively.
On the other hand, the overall SVM Model of accuracy, precision, recall and f-score is 74.06%,
50.74%, 50.86% and 50.57% respectively as tabulated in the above Table 5.4. in which the
highest recognition rate (above 80%) is obtained for ‘ሀ’, ‘ሸ’,’ኘ’, ‘አ’, ‘ከ’ ‘ዐ’, ‘ዠ’, ‘ፐ’ and ‘ቨ’ because
these alphabet signs are a lot in number. The lowest recognition rate is found for ‘ላ’ that has
58.72% accuracy because its shape complex. From the above two experimental results, it has
been shown that SVM based system has given a promising recognition performance.
The classifiers achieve high precision, recall and f- score value for high frequency classes of
Amharic alphabet signs but the low results of those metrics is the cause of the low frequency
alphabet signs.
In general, the classification performance of support vector machine is by far better than neural
network classifier as summarized in Figure 5.4 which shows the comparison of both NN and
SVM model with respect to the four performance metrics namely Accuracy, Precision, Recall
and F-score.
30
20
10
0
Accuracy Precision Recall F-Score
NN SVM
57
According to the above Figure 5.4, a high value of precision in NN and SVM indicates that the
proposed models can correctly classify those samples as alphabet sign which really has alphabet
sign and also the number of selected signs is more relevant. A high recall shows that the many of
the sign data introduced were recognized and also high relevant signs selected while a low recall
shows that the major of the sign cases were ignored. The other high value of f-score indicate that
the best values are recorded at precision and recall performance metrics but the low value is the
cause of worst value at precision and recall.
5.6 Discussion
The proposed Amharic sign language recognizer is evaluated from the capability of learning
machines (SVM and NN) to recognize Amharic sign language into the corresponding text
character. Learning-ability of the selected machines is evaluated by training the 90 % of the
dataset and obtaining their accuracy by using the remaining 10% for testing in each experiment
(fold).
The experimental result of SVM showed significant performance better than NN even though it
uses more memory. The problem of speed and memory may not be a series problem for Amharic
sign language recognition into text character because the fundamental concern is first obtaining a
model that recognizes Amharic sign language with better accuracy. Then speed and memory may
be the second issues when there is an optimization of the classifier. From this perspective, SVM
classifies better than NN regardless of its high memory usage in storage instances.
Past attempts to recognize Amharic sign language achieved fairly good results, but this is mainly
due to the use of few characters. One such attempt is that of Legesse Zerubabel who has attained
the recognition rate of 88.08% and 96.22% using neural network with PCA driven features and
neural network with harr-like features respectively [15]. However, he only focused on ten
selected basic alphabets signs which are ሀ, መ, ሠ, ረ, ሰ, ሸ, በ, ነ, ኘ and አ. In these characters’ hand
shape, there is very few similarities among them. Due to this, his work has scored better
accuracy result. The other related work was done by Abadi Tsegaye [2] who used motion
features (such as angle and direction) and the overall system performance is 71.88% but he only
considered the second, the third, the fourth, the fifth and the sixth orders of signs and also
Abadi’s work only focused on trajectory identification not on recognition. The two mentioned
works used sample data from limited number of signers. In our proposed approach, we covered
58
all the basic and some derived alphabet signs based on shape, motion and color features with NN
and SVM classifiers as it is described earlier. As we can see from Tables 5.2, 5.3 and 5.4 our
proposed system has less accuracy result than previous work that was done by Legesse
Zerubabel. This is attributed due the number of classes used (52 vs 10) and the less similarity of
the selected characters in Legesse’s work.
The answer of the research question which was mentioned in Chapter one is the system correctly
recognizes the basic family and some derived Amharic alphabet signs based on NN and SVM
models. From the two models, SVM is found to be more preferable classifier than NN as the test
result in SVM shows very good performance. It is worth mentioning that our result and findings
also render positive response towards answering the two research equations.
59
Chapter Six: Conclusion and Future Work
6.1 Conclusion
In Ethiopia there are millions of people who are living with hearing problem. These people need
ETHSL as their communication channel among themselves. However, they are living with other
people who communicate with spoken language. The hearing challenged people do not have a
skill to use spoken language well while the hearing people do not have a skill to interpret the
ETHSL. This gap makes the life of hearing impaired people very challenging. Local researches
efforts to fill the gap among those people are very useful. This study is part of the efforts that
were applied to solve the communication problem of hearing impaired people with the hearing
ones.
In this study, an attempt has been made to design and implement a system which is capable of
recognizing Amharic sign language characters. The system has four main parts: Image
preprocessing, Segmentation, Feature extraction and Classification. The first three parts use
various image processing techniques and the last uses two classifiers (NN and SVM) to conduct
the required task.
The system starts by accepting RGB video frame of an Amharic alphabet signs. A video was
captured with a 1280x720 dimension. Therefore, handling each pixel one by one was
computationally expensive and hence before starting the process of segmentation, first the input
video passes through some preprocessing tasks, namely, cropping, gray scale conversion,
contrast adjustment and sharpening. Then the subsequent outcome passes through segmentation
which is implemented by adaptive thresholding to prepare the binary image.
In addition to Amharic alphabet signs segmentation which discriminates the targeted object from
back ground, extraction process that produces different feature vectors was done. The feature
vectors were created by combining shape feature descriptors, motion feature descriptors and
color feature descriptors altogether by applying different feature extraction algorithm which is
Fourier descriptor algorithm.
The last part of the system is sign recognition. For sign classification, NN and SVM classifiers
are employed. Experimental result illustrates that the NN and SVM classifiers achieved an
overall accuracy of 57.82% and 74.06% respectively. The reason for the low overall system
60
performance is the non-existence of well-constructed corpus for experimentation. Thus, we
captured videos of signs from different signers but a number of challenges occurred, including:
Poor video quality, Mobile vibration, and lighting problem. The cause of these problems are due
to the fact that we used junior signers which in return created inconsistency in hand flinging
during signing.
In conclusion, our developed system that recognizes more alphabet signs than the existing or the
previous works for Amharic sign language and it has promising recognition accuracy.
61
performed by different signers with great diversity in background and illumination. This
will significantly help the training of classifiers not to be shallow.
62
References
[1] Masresha Tadesse, "Automatic translations of Amharic text to Ethiopian Sign Language,"
Master's Thesis, Addis Ababa Universty, Addis Ababa, Ethiopia, 2010.
[2] Abady Tsegaye, "Offline candidate hand gesture selection and trajectory determination for
continuous Ethiopian sign language," Master's Thesis, Addis Ababa University, Addis
Ababa, Ethiopia, 2011.
[3] Ethiopian National Association for Deaf, "BERTAT," Yearly Magazine, Ethiopia, 1997.
[4] Najeefa Nikhat Choudhury and Golam Kayas, "Automatic Recognition of Bangla Sign
Language," Bachelor's Project, Department of Computer Science and Engineering,BRAC
University, (Knoxville - USA), 2012.
[5] Daniel Mart´ınez, "Sign Language Translator using Microsoft Kinect XBOX 360TM,"
Department of Electrical Engineering and Computer Science,University of Tennessee, USA,
2012.
[6] Minilik Tesfaye, "Machine Translation Approach to Translate Amharic Text to Ethiopian
Sign Language," Lecturer, Faculty of Informatics, St. Mary’s University College, P.O.Box
18490, Addis Ababa, Ethiopia.
[7] Pradeep Kumar and Manjunatha, "Performance Analysis of KNN, SVM and ANN
Techniques for Gesture Recognition System," Indian Journal of Science and Technology,
vol. 9(51), December 2016.
[8] Nagarajan, Subashini and Balasubramanian, "Visual Interpretation of ASL Finger Spelling
using Hough Transform and Support Vector Machine," International Journal of Advanced
Research in Computer and Communication Engineering , vol. 4, no. 6, June 2015 .
[9] Nagarajan and Subashini, "Static Hand Gesture Recognition for Sign Language Alphabets
using Edge Oriented Histogram and Multi Class SVM," International Journal of Computer
Applications (0975 – 8887), vol. 82, no. 4, November 2013.
[10] Tavari and Deorankar, "Implementation of Neural Network based Hand Gesture
Recognition System," International Journal Of Engineering And Computer Science, vol. 3,
no. 6, 2014.
63
[11] Anand, Sachin and Urabinahatti, "Performance Comparison of Three Different Classifiers
for HCI Using Hand Gestures," International Journal of Modern Engineering Research
(IJMER), vol. 2, no. 4, pp. 2857-2861, July-Aug 2012.
[12] Naidoo, Omlin and Glaser, "Vision-Based Static Hand Gesture Recognition Using Support
Vector Machines," University of the Western Cape, South Africa, on 22 May , 2014.
[13] Yang Quan and Peng Jinye, "Chinese Sign Language Recognition for a Vision-Based Multi-
features Classifier," in International Symposium on Computer Science and Computational
Technology, Shaanxi Xi’an, P. R. China,, 2008.
[14] Atiqur, Ahsan, Ibrahim and Sujit, "Recognition of Static Hand Gestures of Alphabet in
Bangla Sign Language," IOSR Journal of Computer Engineering (IOSRJCE), vol. 8, no. 1,
pp. 7-13, 2012.
[15] Legesse Zerubabel, "Ethiopian Finger Spelling Classification: A Study To Automate
Ethiopian Sign Language," Master's Thesis, Addis Ababa University, Addis Ababa,
Ethiopia, 2008.
[16] Matthews and Baoill, " the Irish Deaf Community (Volume 2): The Structure of the Irish
Sign Language," The Linguistics Institute of Ireland, Dublin, Ireland, 2000.
[17] Samuel Teshome, "Isolated Word-level Ethiopian Sign Language Recognition," Master's
Thesis, Addis Ababa University, Addis Ababa,Ethiopia, 2013.
[18] Shujjat, Gourab, Donald, Serge and Chris, "Sign Language Analysis and recognition: A
Preliminary Investigation," in 24th International Conference on Image and Vision
Computing New Zealand, (IVCNZ 2009).
[19] Fikrte Shiferaw, "Challenges And Opportunities Of Teaching Signed Afan Oromo In Case
Of Sebeta Special Needs Education Teachers’ College," Master's Thesis, Addis Ababa
University, Addis Ababa, Ethiopia, 2014.
[20] Yonas Fantahun and Kumudha Raimond, "Ethiopian Sign Language Recognition Using
Artificial Neural Network," in 10th International Conference on Digital Object
Identifier,PP 995-1000, 2010.
[21] Daniel Zegeye, "Amharic Sentence To Ethiopian Sign Language Translator," Master's
Thesis Addis Ababa University, 2014.
64
[22] Ethiopian Sign Language Community, the deaf people of Ethiopia, people and language
detail report, 2005.
[23] Tefera Gimbi, "Recognition of Isolated Signs in Ethiopian Sign Language," Master's Thesis,
Addis Ababa University, 2014.
[24] Nicolas Pugeault and Richard Bowden, "Spelling It Out: Real-Time ASL Fingerspelling
Recognition," in IEEE Work-shop on Consumer Depth Cameras for Computer Vision,
Barcelona, Spain, 2011.
[25] Ricco and Tomasi , "Fingerspelling Recognition through Classification of Letter-to-Letter
Transitions," in Proceedings of ACCV (3), pp. 214-225, 2009.
[26] Susanna Ricco and Carlo Tomasi, "Finger Spelling Recognition through Classification of L
etter-to-Letter Transitions," In Proceedings of ACC, vol. 3, pp. 214-225, 2010.
[27] Joyeeta Singha, "Indian Sign Language Recognition Using Eigen Value Weighted
Euclidean Distance Based Classification Technique," vol. 4, no. 2, 2013.
[28] Daniel Assefa, "Amharic Speech Training For The Deaf," Master's Thesis, Addis Ababa
University, Addis Ababa ,Ethiopia, 2006.
[29] Alemayehu Teferi, "Ethiopian Sign Language Manual," User Guide on Ethiopian Sign
Language, 2007.
[30] Dr. Anand Singh, "Automatic Recognition of Dynamic Isolated Sign in Video for Indian
Sign Language," Master's Thesis, GLA University, India, 2015.
[31] Christine Fernandez-Maloigne, "Advanced Colour Image Processing and Analysis,"
Springer Science and Business Media New York, 2013.
[32] Tinku Acharya and Ajoy K. Ray, Image Processing Principles and Applications, John Wiley
, 2005.
[33] Chaudhuri, Mandavya, Badela and Gosh, "Optical Character Recognition Systems for
Different Languages with Soft Computing," Springer International Publishing, Studies in
Fuzziness and Soft Computing 352, AG 2017.
[34] Nadia, Albelwi and Alginahi, "Real-Time Arabic Sign Language Recognion," IEEE Paper,
Research Gate, January 2012.
[35] ZHANG Gang, MA Zong-min, NIU Lian-qiang and ZHANG Chun-ming, "Modified
65
Fourier descriptor for shape feature extraction," in Central South University Press and
Springer-Verlag Berlin Heidelberg, 2012.
[36] Direkoglu, Mark and Nixon, "Shape Classification via Image-based Multiscale
Description," Pattern Recognition , ELSVIER, pp. 2134-2146, 2011.
[37] Jonathan Rupe, "Vision-Based Hand Shape Identification for Sign Language Recognition,"
Master's Thesis, Rochester Institute of Technolog, 2005.
[38] Tahir, Hussain, Samad, Husain and Rahman, "Human Shape Recognition using Fourier
Descriptor," Journal of Electrical and Electronic Systems Research, vol. 2, June 2009.
[39] Corneliu Lungociu, "Real Time Sign Language recognition Using Artificial Neural
Networks," Studia Univ,Babes-Bolyai,Informatica, vol. LVI, no. 4, 2011.
[40] Amit Sharan, "Character Recognition Using Fourier Coefficients," Master's Thesis ,in Texas
Tech University, December 1993.
[41] Klimis Symeonidis, "Hand Gesture Recognition Using Neural Networks," Master's Thesis,
UniS, On August 23, 2000.
[42] Sanjay Meena, "A Study on Hand Gesture Recognition Technique," Master's Thesis,
National Institute Of Technology, Rourkela, India, 2011.
[43] Kenji Suzuki, Artificial Neural Networks-Methodological Advances And Biomedical
Applications, In Tech Book, Ivana Lorkovic, 2011.
[44] Coolen, Kühn and Sollich, Theory of Neural Information Processing Systems, New York:
Oxford University Press, 2005.
[45] George Awad, "A Framework for Sign Language Recognition using Support Vector
Machines and Active Learning for Skin Segmentation and Boosted Temporal Sub-units,"
Master's Thesis Dublin City University(DCU), in 2007.
[46] D.Karthikeyan and Mrs.G.Muthulakshmi, "English Letters Finger Spelling Sign Language
Recognition System," International Journal of Engineering Trends and Technology
(IJETT), M.S.University,Tirunelveli,India, vol. 10 , no. 7, Apr 2014.
[47] Nikhil, Shreyas, Sumesh, Gowranga and Bhakthavathsalam, "Implementation and
Comparison of Machine Learning Algorithms for Recognition of Fingerspelling in Indian
Sign Language," Society For Science And Education United Kingdom, vol. 5, no. 5, 20th
66
August, 2017.
[48] Marijana, Sanja and Natasa, "A Comparison of Machine Learning Methods in a High-
Dimensional Classification Problem," Business Systems Research, DE GRUYTER, vol. 5,
no. 3, September 2014.
[49] Azher Uddin and Shayhan Ameen, "Hand Sign Language Recognition for Bangla Alphabet
using Support Vector Machine," IEEE International Computer Science and Engineering
Conference, 2016.
[50] Michael Kearns and Dana Ron, "Algorithmic Stability and Sanity-Check Bounds for Leave-
One-Out Cross-Validation," AT&T Labs Research, Murray Hill,New Jersey, January 1997.
[51] Walter Daelemans and Antal van den Bosch, "Memory-Based Language Processing,"
Cambridge University Press 978-0-521-80890-3, 2009.
67
Appendixes
Appendix A: Sample Data Used for System Design
68
69
Appendix B: MATLAB Code
I. Segmentation
%Read from video source
ff= fullPathname;
obj = VideoReader(ff);
nframes = get(obj, 'NumberOfFrames');
%Take sample points for hue difference measure
LF=read(obj, nframes);
FF=read(obj, 1);
[w,hi]=size(FF);
SZ=floor(nframes/2);
[L,no]=bwlabel(BASE_FRAME,8);
BASE_FRAME=dispLabel(L,BF_IDX, BASE_FRAME);
BASE_FRAME=logical(BASE_FRAME);
SE=strel('disk',20);
BASE_FRAME=imerode(BASE_FRAME,SE) & BASE_FRAME ;
BASE_FRAME=bwareaopen(BASE_FRAME, 2000);
SE=strel('disk',30);
BASE_FRAME=imdilate(BASE_FRAME,SE) & BASE_FRAME ;
Return BASE_FRAME
70
if
(( h(i,j) >= .009 && h(i,j) <= .09 ) && s(i,j) >= .23)
Sframe(i,j)=1;
end
end
end
SE=strel('diamond',10);
Sframe=imdilate(Sframe,SE) & Sframe ;
Sframe=bwareaopen(Sframe,1500);
Sframe=imfill(Sframe,'holes');
stats = regionprops(Sframe, {'Centroid','Area'});
areaArray = [stats.Area];
[junk,idx] = max(areaArray);
c = stats(idx).Centroid;
c = floor(fliplr(c));
xcor(index)=c(2);
ycor(index)=c(1);
if k==1
BASE_FRAME=Sframe;
BF_IDX=idx;
end
index=index + 1;
END_FRAME=Sframe;
end
%End function
end
[h0,s0,v0]=rgb2hsv(LF);
[h,s,v]=rgb2hsv(FF);
len=numel(xcor);
coldiff=sqrt(abs(h(ycor(len), xcor( len )) - h0(ycor(1), xcor(1)
)) + abs(s(ycor(len), xcor(len))-s0(ycor(1), xcor(1))) +
abs(v(ycor(len),xcor(len))-v0(ycor(1),xcor(1))));
72
movType=2;%move right
elseif((xcor(count)-xcor(1)) <0 && (ycor(count)-ycor(1)) >0
&& d2 > d1)
movType=2;%move right
elseif((xcor(count)-xcor(1)) >0 && (ycor(count)-ycor(1)))
movType=1;%move left
end
VFS1(1)=movType;
VFS1(2)=sumteta;
VFS1(3)=coldiff;
VFS2(32)=VFS1(1);
VFS2(33)=VFS1(2);
VFS2(34)=VFS1(3);
VFS=VFS2;
end
III. Recognition
%Get extracted video features
ExtractedVFS=handles.ExtractedVFS;
valNN={};
valSVM={};
input=[];
output=[];
73
%Load the dataset
DS=load('dataset.mat');
input=DS.input;
input=cell2mat(input');
output=DS.output;
74
nndisp=strnn;
else
nndisp=str{1,strnn};
end
if isempty(str{1,strsvm})
svmdisp=strsvm;
else
svmdisp=str{1,strsvm};
end
set(handles.editresult, 'string',nndisp);
set(handles.edit4, 'string',svmdisp);
75
76
77