OFF-LINE YORUBA HANDWRITING SENTENCE RECOGNITION
USING CHERIET RECOGNITION ALGORITHM
BY
YAHAYA ABDULLAHI
Department of Computer Science
Kwara State University, Malete
and
JUMOKE F. AJAO
Department of Computer Science
Kwara State University, Malete
MAY, 2017
Abstract:
Handwriting is the human way of communicating using written media.
Nowadays, there are a lot of changes of technology in terms of communication.
Handwriting offers an attractive and efficient method to interact with computer, in
order to enhance human computer interaction. This research work focus on Yoruba
handwriting sentence recognition using Cheriet algorithm. Many handwriting
recognition system have been developed but these only captured Yoruba
handwriting word, there is need to develop Yoruba handwriting sentence
recognition system.
Four methods were adopted for the recognition process as follows: Data
acquisition, image processing, feature extraction and recognition. Fifty
handwriting were acquired from literate indigenous writers, which was subjected
to image preprocessing to enhance the quality of the digitized image. The features
of the pre-processed image were extracted using Surf Algorithm and the extracted
feature vectors were subjected to Cheriet algorithm for recognition. The
recognition system developed was evaluated based on the recognition rate and
92.8% recognition rate was achieved.
1
I.
INTRODUCTION
Handwriting is the human way of communicating using written media.
Nowadays, there are a lot of changes of technology in terms of communication.
Handwriting offers an attractive and efficient method to interact with computer,
such as a tool developed at this time that is able to receive input in the form of
handwriting data. The tool also requires a method to recognize input in the form of
handwriting data. Research on handwriting recognition issue becomes important to
provide solutions to some of these problems. There is difference in the writing
patterns of individuals; so computers need to recognize the different of handwriting
patterns of individuals regardless of how it is written (Lulu, Widodo and Cipta
2014).
Handwriting recognition can be classified as e i t h e r off-line o r o n -line.
In off-line recognition, the writing is usually captured optically by a scanner and
the completed writing is available as an image. But, in the on-line system the
two dimensional coordinates of successive points a r e represented as a function
of time and the order of strokes made by the writer are also available (Anita,
2012). Several applications including mail sorting, bank processing, document
reading and postal address recognition require off-line handwriting recognition
systems. As a result, the off-line handwriting recognition continues to be an
active area for research towards exploring the newer techniques that would
improve recognition accuracy.
This Research work used an off-line system; the technique to be adopted
yielded comparably high recognition accuracy levels. Although many highly
accurate systems have been developed to recognize handwritten numerals and
characters but such system has not been extended to Yoruba Handwriting sentence
recognition. This has been ascribed to the difficult nature of Yoruba handwriting,
including the diversity of character patterns, ambiguity and illegibility of
characters, and the overlapping nature of many characters in Yoruba word due to
the presence of diacritic sign (Gader, 2013). This research attempts to evaluate the
2
effect of recognition of Yoruba Handwriting Sentence through the use of Cheriet
Algorithm.
Statement of the Problem
Many Yoruba Handwritten recognition system developed catered for
Yoruba characters and Yoruba words recognition. There is need to extend the
recognition to Yoruba sentence. The research work will focus on recognition of
Yoruba sentence using Cheriet algorithm.
Aim of the Study
The aim of this research work is to develop a system that can recognize
Yoruba handwriting sentence using Cheriet Algorithm.
Objectives of the Study
i.
To acquire Yoruba sentence to be recognize
ii.
To design a recognition system for the Yoruba sentence.
iii.
To implement the system design in ii using Java programming language
iv.
To evaluate the recognition system using the recognition rate
Significance of the Study
In order to save the indigenous from extinction, it is very important to
incorporate Yoruba recognition system to information technology
Scope of the Study
This research focused on the offline handwritten technology in which
samples are collected from learned Yoruba writers. The system correct any mistake
i.e. irrelevant information in handwritten word due to user dependency. The system
was developed as a standalone such that anyone can install and use it on a personal
computer (PC).
3
II.
Yoruba Language
Yoruba is a language spoken in West Africa, mainly in Nigeria. The
number of speakers of Yoruba is approaching 30 million (Mikael, 2007). The
Yoruba dialect consists of several dialects. The various Yoruba dialects in the
Yoruba land of Nigeria can be classified into five major dialect areas: Northwest,
Northeast, Central, Southwest and Southeast (Adetugbọ, 1982). Clear boundaries
cannot be drawn, peripheral areas of dialectal regions often having some
similarities to adjoining dialects.
The pronunciation of the letters without diacritics corresponds more or less
to their International Phonetic Alphabet equivalents, except for the labialvelar stops [k͡p] and [ɡ͡b], in which both consonants are pronounced simultaneously
rather than sequentially. The diacritic underneath vowels indicates an open vowel,
pronounced with the root of the tongue retracted (so “ẹ” is pronounced [ɛ̙] and ọ
is [ɔ̙]). “ṣ” represents a post alveolar consonant [ʃ] like the English “sh”, “y”
represents a palatal approximant like English y, and j a voiced palatal plosive [ɟ],
as is common in many African orthographies.
Yoruba alphabet can be broken down into three tiers:
i.
Tonal Tier
ii.
The Under dot Tier
iii.
The Character Tier and The diagraph
4
A B D E Ẹ F G Gb H I J K L M N O Ọ P R S Ṣ T U W Y
a b d e ẹ F g gb H i
j k l m n o ọ p r S ṣ t u w y
Figure 2.1: The current orthography of Yoruba Alphabets without
diacritical marks
Á À Ā É È Ē
Ẹ/
Ẹ́ /
Ẹ̀ /
Ẹ̄ /
E̩
É̩
È̩
Ē̩
Ọ/
Ọ́/
Ọ̀ /
Ọ̄ /
O̩
Ó̩
Ò̩
Ō̩
Í Ì
Ī Ó
Ò Ō
á à ā é è Ē ẹ / e̩ ẹ́ / é̩ ẹ̀ / è̩ ẹ̄ / ē̩ í ì
ī ó
ò ō ọ / o̩
Ṣ/
S̩
ọ́ / ó̩ ọ̀ / ò̩ ọ̄ / ō̩ ú ù ū ṣ / s̩
Figure 2.2: The current orthography of Yoruba Alphabets
with diacritical marks
5
Ú Ù Ū
III
Literature Review
Researchers have utilized many different approaches for both the
recognition and segmentation tasks of word recognition. Some researchers have
used conventional, heuristic techniques for both sentence segmentation and
recognition.
Jumoke F., Stephen O. Elijah O. and Odetunji O. makes use of the
Information Theory Approach to solve Yoruba Handwriting Word Recognition
Quality Evaluation of Pre-processing Attributes, an Entropy preprocessing stage
measure closer to the entropy measure of the original Information was obtained.
Hong Lee, and Brijesh Verma uses the approach of Binary segmentation algorithm
for Segmentation of English cursive handwriting recognition and achieved a
recognition rate of 83.1%. Ajao JF, Jimoh RG, and Olabiyisi uses the artificial
neural network for handwritten recognition system of destination address on
envelopes and achieved a recognition rate of 96%.
Blumenstein and Verma (2013) uses the Neuro-heuristic algorithm to
conduct a research for varieties of New Segmentation Algorithm for Handwritten
Word Recognition and obtained a results of up to 76.52% for a test set of
segmentation point patterns. Femwa O. conducted a research on Development of a
writer-independent online handwritten character recognition using modified hybrid
neural network (geometrical and statistical features) and obtained a recognition rate
of 96%. Vijay Laxmi Sahu and Babita Kubde work on Offline Handwritten
Character Recognition Techniques using Neural Network and achieved an accuracy
of 97%.
6
METHODOLOGY
Data Acquisition
Data were acquired from adult indigenous writers. The recognition of the
Yoruba sentence was computed. The approach is aimed at recognizing this Yoruba
sentence using the Cheriet algorithm after passing through the preprocessing stages.
This is to provide information required in the preprocessing stage as a preparatory
level for the handwritten Yoruba recognition system to be developed. Samples of
Yoruba words in handwritten collected from indigenous writers is presented in
figure 3.1
This research uses the concept of word recognition with the use of Cheriet
Algorithm where the characters are recognized using pre-processed character
database. Figure 3.2 shows the system flow chat for Yoruba sentence recognition.
7
Figure 3. 1 Samples of Yoruba sentence collected
START
LOAD SENTENCE FROM
IMAGE FOLDER
CONVERT
IMAGE TO GRAY
SCALE
PREPROCESSING
BINARIZE GRAY
SCALE IMAGE
FEATURE
EXTRACTION
NOT RECOGNIZED
RECOGNITION
NO /
YES
RECOGNIZED
8
DISPLAY SENTENCE
OPEN AS
NOTEPAD
OPEN AS
MS-WORD
STOP
Figure 3. 2: System Flowchart
9
Pre-Processing
The aim of preprocessing stage is the removal of all elements in the word
image that are not useful for recognition process (Benouareth, 2008). The generic
image pre-processing involves the following series of stages that are carried out in
order to extract important information from an image and thereby reducing
unwanted information from the image so as to enhance the recognition process.
The preprocessing carried on the Yoruba sentence recognition are as follows:
i.Conversion of RGB to Grayscale: Conversion of RGB to Grayscale coverts
image scanned into the computer to Grayscale image to enhance detail information
needed from the image. Given that RGB value of a color is (R, G, B) such that R,
G, and B are integers between 0 and 255. The grayscale weighted average, X, is
given by the formula:
X = 0.299R + 0.587G + 0.114B
(1)
ii.Binarization: Binarization is the transformation of a grayscale image into a black
and white image through thresholding inspired by Niblack’s algorithm. It calculates
a pixel-wise threshold by sliding a rectangular window over the gray level image
given by the equation 1:
TNiblack = m + k * s
TNiblack = m + k √
1
𝑁𝑃
TNiblack = m + k √
(2)
∑(𝑝𝑖 − 𝑚)2
∑ 𝑃𝑖 2
𝑁𝑃
− 𝑚 2 = 𝑚 + 𝑘 √𝐵
(3)
(4)
Where NP is the number of pixels in the gray image, m is the average value of the
pixels pi, and k is a fixed constant.
10
Niblack ALGORITHM
import os, sys
from PIL import Image
inFile = ''
outFile = ''
if len(sys.argv) != 3:
print 'Input format error!'
else:
inFile = sys.argv[1]
outFile = sys.argv[2]
im = Image.open(inFile).convert('L')
for i in range(im.size[0]):
for j in range(im.size[1]):
if im.getpixel((i,j)) > 127:
im.putpixel((i,j), 255)
else:
im.putpixel((i,j), 0)
im.show()
im.save(outFile)
Proposed Recognition Algorithm Method
The original grey scaled image is binarized using Cheriet algorithm by
selecting automatically a threshold value for a given image. Finally, the image is
converted to skeleton format to allow users verity of writing device, pen tilt and to
suppress extra data. The proposed recognition algorithm is explained in the figure
3.2.
Step 1. Take sentence image from the database.
Step 2. Perform pre-processing on taking image.
Step 2.1 convert image sentence to gray scale
Step 2.2 convert gray scale image to binarization using the Niblack
approach
Step 3. Perform recognition of sentence
11
Due to the simplicity of the proposed recognition technique, it is very fast
and performs well in most of the cases. But has a limitation in it recognition of
some characters with diacritical marks such as m, n, u, v and w in which over
segmentation occurs and this technique fails to find accurate character boundaries.
Cheriet Algorithm
In this algorithm, the words is first decomposed in two parts; main body of
letters and secondary parts (the diacritics parts) by contour following.
The primary parts undergoes a horizontal decomposition distinguishing two areas:
median and overflowing zones as shown in figure 3.3.
The overflowing zones contain ascenders and descenders, while the median
contain valleys. Valleys are the horizontal segments connecting adjacent peaks. It
can be seen from the figures that Cheriet algorithm is good except for few
characters with under dot, where over segmentation occurs. Therefore it is required
to integrate this technique with some intelligent method to increase its performance.
It is mention worthy that over segmentation is minimum and occurs for few
characters only. Hence it lessened burden of the algorithm used and therefore
processing speed increased.
1-a)
Original Image
1-b)
2-b)
2-a)
Exterior Contours
Overlapping Zones
Exterior Contours
Figure 3. 3: 1) Exterior contour extraction
12
2) Decomposition in median and
overflowing zones
Word Extraction
This stage aims to remove the white spaces around each word and extract
the effective image size by moving through the black pixels and when it finds white
space more than 100 pixels that means the word is finished and it starts another
word. The main goal of feature extraction is to reduce the data dimensionality and
properly represent the original data in feature space. Features useful for
classification process can be simple features like RGB values in color images, or
complex features like energies from the Fourier Transform or Wavelet Transform
of a time series. The feature extraction process usually consists of three steps
1.
Feature construction is the step in which features are constructed from linear
or non-linear combination of raw features.
2.
feature selection process is done using techniques like relevancy ranking of
individual features
3.
Feature reduction process is used to reduce the no. of features especially
when too many features are selected compared to the no. of feature vectors. These
three steps are not mandatory in the feature extraction process.
13
Algorithm for Surf
do
do
do
for all interest area in given input image,
calculate Hessian Matrix H (5×5)
end
Identify two interest area with same determinant value;
Mark as a feature;
end
divide the feature (interest area) into 4×4 subarea;
find deviation in x and y axis (estimating transformation);
get the angle of deviation as a trace of Hessian matrix;
recover the original image by inverse transformation;
end
14
The SURF Algorithm
1.
Detection
Automatically identify interesting features
Where
𝐻 (𝑋, 𝜎) = [
𝐿𝑥𝑥 (𝑋, 𝜎)
𝐿𝑥𝑦 (𝑋, 𝜎)
𝐿𝑋𝑦 (𝑋, 𝜎)
]
𝐿𝑦𝑦 (𝑋, 𝜎)
(5)
𝜕2
𝑔(𝜎)
𝜕𝑥 2
𝜕2
𝐿𝑥𝑦 (𝑋, 𝜎) = I(X) ∗
𝑔(𝜎)
𝜕𝑥 2
𝐿𝑥𝑥 (𝑋, 𝜎) = I(X) ∗
2.
Description
Each interest point should have a unique description that does not depend
on the features scale and rotation.
𝑣 = {∑ 𝑑𝑥, ∑|𝑑𝑥|, ∑ 𝑑𝑦, ∑|𝑑𝑦|}
(6)
3.
Matching
Given and input image, determine which objects it contains, and possibly a
transformation of the object, based on predetermined interest points.
∇2 𝐿 = 𝑡𝑟(ℋ) = 𝐿𝑥𝑥 (𝑋, 𝜎) + 𝐿𝑦𝑦 (𝑋, 𝜎)
15
(7)
Performance Evaluation of the System
The performance of the system is measured by the recognition rate achieved
by the developed system. The Cheriet algorithm used in the development of the
system will be evaluated by performing a recognition test in finding the percentage
of sentence recognized by the system and those not recognized. The performance
metrics that will be used for the testing is given below:
The over-all Recognition rate is given thus,
∑ 𝑅𝐸
R.R = ∑
𝑇𝑆
× 100
(8)
Where 𝑅𝐸 is recognized sample and 𝑇𝑆 is tested sample, and R.R is Recognition
Rate. The over-all Not-Recognition rate is given as:
∑ 𝑁𝑇
N.R = ∑
𝑇𝑆
× 100
(9)
Where 𝑁𝑇 is not-recognized sample and 𝑇𝑆 is tested sample
16
RESULT AND DISCUSSION
Testing and Result
Data were acquired from adult indigenous writers. The recognition of the
Yoruba sentence was computed. The approach is aimed at recognizing this Yoruba
sentence using the Cheriet algorithm after passing through the preprocessing stages.
This is to provide information required in the preprocessing stage as a preparatory
level for the handwritten Yoruba recognition system to be developed. Table 4.1 and
4.2 shows the detailed description of samples gathered, tested and the results
obtained from each test.
Discussion of Results
To this end, this research project has been able to achieve it intended
purpose based on the result obtained on it usability. Many researchers mention
recognition and segmentation as a part of their overall systems, however few report
their findings at the segmentation level. This research has focused on this very
important area and has produced commendable results which can easily be
compared to other researchers in the field.
17
Author
Problem Solved
Technique
Method
Sample
Collected
Result
Obtained
Blumenstein
and Verma
varieties of New
Segmentation
Algorithm for
Handwritten
Word
Recognition
Neuroheuristic
algorithm
set of
segmentatio
n point
patterns.
obtained a
results of
up to
76.52%
Jumoke F.,
Stephen O.,
Elijah O.,
and Oladayo
O.
Offline Yoruba
Handwritten
Word
Recognition
Hidden
Markov Model
Approach
Yoruba
written
word was
gathered
achieved a
recognition
rate of
95.6%
Han and
Sethi
Segmentation of
words on 50
envelopes from
real mail pieces
Achieved
an
accuracy of
85.7%
Eastwood et
al.
recognition of
handwriting
from CEDAR
segmentati
on of
words on
50
envelopes
Sample
Handwritte
n gathered
from CDROM
Yahaya
Abdullahi
Off-Line
recognition of
Yoruba
Sentence
heuristic
algorithm
Otsu algorithm
Using Cheriet
Algorithm
18
50 written
Yoruba
sentence
Achieved
an
accuracy of
75.9%
Achieved
an
accuracy of
92.8%
Table 4. 1: sentence gathered and result from using Cheriet Algorithm
Tested
Samples
Sample
Recognized
Not
R.R
N.R
Recognized
(%)
(%)
Gbogbo Okunrin ati Obirin
50
48
2
96
4
Ifemi otito ni ajoke
50
49
1
98
2
Baba ati iya
50
50
0
100
0
Nibolowa lati ana
50
45
5
90
10
50
40
10
80
20
𝑜́ 𝑠𝑖̀ 𝑦𝑒 𝑘𝑖̇ 𝑤𝑜𝑛 𝑜́ 𝑚𝑎́ 𝑎 ℎ𝑢̀ 𝑤𝑎̀ 𝑠𝑖́
Table 4. 2: Sample of sentence gathered
Handwritten Image
Ifemi otito ni ajoke
Gbogbo Okunrin ati Obirin
Baba ati iya
Nibolowa lati ana
𝑜́ 𝑠𝑖̀ 𝑦𝑒 𝑘𝑖̇ 𝑤𝑜𝑛 𝑜́ 𝑚𝑎́ 𝑎 ℎ𝑢̀ 𝑤𝑎̀ 𝑠𝑖́
19
III.
SUMMARY CONCLUSION AND RECOMMENDATIONS
Summary
The objective of image preprocessing in handwriting recognition is to
ensure original image restoration. In the present work, a recognition method is
devised to adapt the information theory and related techniques in the development
of a robust and accurate Yoruba sentence recognition system. Considering all the
aspects discussed in the previous sections, the next steps to provide better Yoruba
handwriting sentence recognition systems are obvious. The recognition module will
be able to determine the preprocessing stages that are required to be performed on
the samples handwriting. This will enhance the performance of the Yoruba
recognition to be developed.
Conclusion
An intelligent recognition algorithm has been presented in this research,
producing good results. It was used to recognize different printed handwritten
words gathered from different indigenous writers. With some modifications, more
testing shall be conducted to allow the technique to be used as part of a larger
system. It is therefore hoped that further research can be dedicated to analyzing and
improving the results of this very important procedure.
Recommendations
In future work, the recognition technique will be improved in a number of
ways. As a result, under recognition was noticeable in some words. Therefore, the
algorithm shall be modified so that it will be possible to detect a smaller number of
incorrect recognition points, while at the same time recovering more correct
recognition. This can be achieved by looking for more features or possibly
enhancing the current feature detection methods. More sentence shall also be used
in training and testing, and finally the technique shall be integrated into a complete
handwriting recognition system.
20
IV.
REFERENCES
Anita Jindal, Renu Dhir and Rajneesh Rani, (2012) “Diagonal Features and SVM
Classifier for Handwritten Gurumukhi Character Recognition,” Volume 2,
Issue 5, ISSN: 2277 128X International Journal of Advanced Research in
Computer Science and Software Engineering.
Asworo E., (2003) “Compatration Between Kohonen Neural Network and
Learning Vector Quantization Methods on Real Time Handwritting
Recognition System”, Institute of Technology Surabaya.
Bamgbose, A., (1976) “Yoruba Orthography,” Ibadan University Press, pp. 15-27.
Bamgboṣe, Ayọ (1965). Yoruba Orthography. Ibadan: Ibadan University Press.
Benouareth A., Ennaji I. and M. Sellami, (2008) “Arabic Handwritten Word
Recognition Using HMMs with Explicit State Duration," EURASIP Journal
on Advances in Signal Processing, Volume, Article ID 247354.
Bozinovic R. M., and Srihari S. N. (1989), “Off-Line Cursive Script Word
Recognition”, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol.
11, pp. 68-83.
Brunelli R. (2009), Template Matching Techniques in Computer Vision: Theory
and Practice, Wiley, ISBN 978-0-470-51706-2.
Casey R. G., and Lecolinet E. (1996). “A survey of Methods and Strategies in
Character Segmentation”, IEEE Trans Pattern Anal Mach Intell, 18 (7), pp.
690–706.
Cheriet M., Suen C. Y., Legault R., Nadal C. and Lam L., (1993) “Building a New
Generation of Handwriting Recognition Systems”, Pattern Recognition
Letters, Vol. 14, 1993, pp. 305-315.
21
Gader P. D. (1996), “Fusion of Handwritten Word Classifiers”, Pattern Recognition
Letters, Vol. 17, 1996, pp. 577-584.
Gader P., Whalen, M., Ganzberger, M., Hepp, D., “Handprinted Word Recognition
on a NIST Data Set”, Machine Vision Applications, Vol. 8, 1995, pp 31-40.
Ibrahim A and Odejobi O., (2016) “A System for the Recognition of handwritten
Yoruba characters.” Obafemi Awolowo University, Ile-Ife, Nigeria. AGIS.
Jumoke F. Ajao, Stephen O. Olabiyisi, Elijah O. Omidiora and Oladayo O.
Okediran, (2012) “Hidden Markov Model Approach for Online Yoruba
Handwritten Word Recognition” British Journal of Mathematics &
Computer Science 18(6): 1-20, , Article no.BJMCS.27567 ISSN: 22310851
Lulu C. Munggaran, Suryarini Widodo, and Cipta A.M, (2014), “Handwritten
Pattern Recognition Using Kohonen Neural Network Based on Pixel
Character” (IJACSA) International Journal of Advanced Computer Science
and Applications, Vol. 5, No. 11.
Martin G. L., Rashid M., and Pittman J. A., (1993) “Integrated Segmentation and
Recognition through Exhaustive Scans or Learned Saccadic Jumps”, Int’l J.
Pattern Recognition and Artificial Intelligence, Vol. 7, pp. 831-847.
Mikael Parkvall, Världens Marhsall, and största spark, (2007)" (The World's 100
Largest Languages in 2007), in National encyclopedia.
Mineichi Kudo; Jack Sklansky (2000). "Comparison of algorithms that select
features
for
pattern
classifiers". Pattern
Recognition. 33 (1):
25–
41. doi:10.1016/S0031-3203(99)00041-2
Omnigot A., (2013), “The Yoruba alphabets and it pronunciation,” Accessed at
http://www.omniglot.com/writing/yoruba.htm And Support Vector
22