FREEMAN CODE BASED ONLINE HANDWRITTEN CHARACTER RECOGNITION FOR MALAYALAM USING BACKPROPAGATION NEURAL NETWORKS

Advanced Computing: An International Journal ( ACIJ ), Vol.3, No.4, July 2012
DOI : 10.5121/acij.2012.3407 51
FREEMAN CODE BASED ONLINE HANDWRITTEN
CHARACTER RECOGNITION FOR MALAYALAM
USING BACKPROPAGATION NEURAL NETWORKS
Amritha Sampath1
, Tripti C2
and Govindaru V3
1
Department of Computer Science and Engineering, Rajagiri School of Engineering and
Technology, Kochi, India
amrithasampath@yahoo.com
2
Department of Computer Science and Engineering, Rajagiri School of Engineering and
Technology, Kochi, India
triptic@rajagiritech.ac.in
3
Computational Linguistics, Centre for Development of Imaging Technology, Kerala,
India
neithalloor@gmail.com
ABSTRACT
Handwritten character recognition is conversion of handwritten text to machine readable and editable
form. Online character recognition deals with live conversion of characters. Malayalam is a language
spoken by millions of people in the state of Kerala and the union territories of Lakshadweep and
Pondicherry in India. It is written mostly in clockwise direction and consists of loops and curves. The
method aims at training a simple neural network with three layers using backpropagation algorithm.
Freeman codes are used to represent each character as feature vector. These feature vectors act as inputs
to the network during the training and testing phases of the neural network. The output is the character
expressed in the Unicode format.
KEYWORDS
Freeman code;Backpropagation Neural Networks; Unicode
1. INTRODUCTION
Optical character recognition (OCR) can be based on conversion of typewritten or printed
characters as in textbooks or it can deal with conversion of handwritten text into machine
editable form. Both have their own applications. Conversion of handwritten characters is
important for making several important documents related to our history, such as manuscripts,
into machine editable form so that it can be easily accessed and preserved. A search is difficult
when information is available in a form which is not recognizable by the machine. If it is
converted into a machine recognizable form, the search becomes fast and easier. The method of
conversion of already existing information is called Offline character recognition. Such systems
are called OCR systems.
Online recognition is important as an alternate method of data input. Languages like Malayalam
have large character set, hence difficult to have a keyboard which can be used easily. So for
online data acquisition a digital pen or stylus can be used. But due to limitations of the device or
speed of writing, or tremble, it is possible for a single character to be broken into different parts,
hence creating confusion in recognition.
Handwritten character recognition is generally more difficult than conversion of printer and
typed characters since, in latter there is a standard set of fonts to which it can be mapped.

52
However, handwritings vary from person to person and also for a person it may vary from time
to time according to his/her mood, urgency, etc. Hence handwritten character recognition is a
difficult task. Existing character recognition software in Malayalam focuses on conversion of
printed or typed texts.
Handwritten character recognition has been developed for several languages. But its difficulty
in Malayalam can be attributed to several reasons like complexity and similarity in the way
characters are written and also due to large character set. Malayalam can be written either in the
old lipi or in new lipi as shown in Table 1. Hence number of characters to be recognized will
almost be doubled.
Table 1. Characters in old and new lipis.
Online and offline character recognition requires basically four steps.
1. Pre-processing
2. Feature extraction
3. Recognition
4. Post processing
The method used in these basic steps varies according to the application.
2. RELATED WORK
Though there has been a lot of study on the handwritten character recognition in many
languages, an efficient system in Malayalam has not yet been developed. Most of the research
has been based on the offline character recognition and on typed text. Malayalam consists of
characters with loops and curves, with most of the characters being written in the clockwise
direction.
An OCR system for Malayalam has been developed which uses the number of horizontal and
vertical lines for the identification of the characters[1]. It includes pre-processing, character
extraction and skeltonization phases before the actual recognition takes place. The recognition
module include functions which calculate the number and position of horizontal and vertical
lines which forms the feature that distinguishes each character from another. Offline recognition
of Malayalam characters using chain code histogram and normalised chain code histogram has
also been developed[2]. Chain code is used to represent the boundary of the character and is
stored as location and direction of line segments of specified length. Centroid of the image is

53
also taken to improve the result. Online system which uses a combination of context bitmap and
normalised (x,y) co-ordinates has also been developed[3]. It uses Kohonen network for
recognition. A recognition system developed in Tamil[4], which is another prominent language
used in south India, uses a post-processing stage to distinguish between two confusing
characters.
3. METHOD OF IMPLEMENTATION
The method proposes different processing techniques for each of the four steps mentioned.
3.1. Pre-processing
Pre-processing includes noise removal.
A noise is a mark made on the writing surface which is not to be taken as a part of the input.
Noise will be different from the actual input in its characteristics. A stroke can be defined as a
set of points taken from a pen-down position to pen-up position. It is a trajectory followed by
the pin tip from the point when it makes the first contact with the writing surface to the point
when it leaves the surface. The time taken to make a noise stroke will be either too high or too
low when compared with the average time to make a stoke of the actual character. Also, if noise
is much away from the rectangular area and number of pixels is less than a threshold, it can be
removed as noise.
3.2. Feature Extraction
Feature extraction is the next step after pre-processing. We need to identify unique features that
can be used to uniquely identify every character in the character set of the language.
Feature extracted can be either low level or high level. Low level features include width, height,
curliness, aspect ratio etc of the character. These alone cannot be used to distinguish one
character from another in the character set of the language. So, there are a number of other high
level features which include number and position of loops, straight lines, head lines, curves etc.
One feature that can be used for identification is direction information which is collected online.
It is based on Freeman codes as shown in Figure 1.
Figure 1. Freeman codes[5]
Starting from the point when first contact is made with the writing surface, direction in which
the pen tip moves is recorded. 1 for NE, 2 for E, 3 for SE etc will be stored as a single
dimensional array. Direction is recorded only when there is change in direction to avoid

54
dependence of length of line segments in the character. Also in order to mark crossings of line
segments, a character 9 can be used. This array is used as a feature vector for classification.
Figure 2. Sample input given during training
Figure 2 shows a sample input ‘ga’ given as input for training. The input will be coded into
feature list which is stored as a linked list. For the given sample input, the feature vector is ‘[3,
4, 3, 2, 3, 2, 1, 2, 1, 0, 1, 0, 1, 0, 1, 2, 3, 4, 3, 4, 5]’. An issue, that arise when creating the
feature vector based on direction of pen movement is that, instead of storing a ‘1’ in the feature
vector for the NE direction, it may store it as ‘2’ followed by ‘0’. This issue arises due to
irregularities in writing caused due to the inexperience of the user in using the device, shivering
during writing etc. This can be avoided by extracting the direction formed between points 2
pixels apart rather than adjacent pixels. This greatly helps to reduce the size of feature vector
and makes it more accurate.
3.3. Classification
Several techniques such as k-Nearest Neighbor (k-NN) [6], Bayes Classifier, Neural Networks
(NN), Hidden Markov Models (HMM), Support Vector Machines (SVM), etc exist for the
purpose of classification. One of the commonly used techniques is neural networks. Neural
networks consist of a number of nodes and links arranged as different layers as shown in Figure
3. Different links which connect different layers are associated with weights. Input to a node is
the sum of product of activation and weight associated with the link. The weights must be
selected such that inputs map to their corresponding outputs.
Usually a neural network consists of a training phase, a validating phase and a testing phase.
During the training phase, features extracted will be used for training the network to map input
to output. The training set consists of feature vectors for each of the characters that must be
recognized by the network. A training cycle consists of a forward pass and a reverse pass.
Backpropagation algorithm is a supervised training algorithm which is used to train the network
by adjusting the weights according to Delta rule.
Initially, the weights in the neural network are assigned small random weights. The input is
applied to the network. In each layer, the input is multiplied with the corresponding weights and

55
an activation function such as sigmoid function is applied at each node. This acts like a
squashing function.
The formula of sigmoid activation is: f(x) =1/(1 + e−input
).
The output obtained from the last layer is compared with the expected output. This gives the
error. This is propagated backwards and the weights associated with the links in each layer is
modified as weight(old) + learning rate * output error * output(neurons i) * output(neurons i+1)
* ( 1 - output(neurons i+1) )
Figure 3. A simple neural network
Figure 4. A node in a neural network[7]
After one training cycle is complete, the next input feature vector is applied and the process is
repeated for all the feature vectors in the training set. This completes one training epoch. The

56
network requires several training epochs before the network learns to recognize the characters
in the training set.
After training, the input feature vectors are applied to test the network. In the validating phase,
feature vectors that were not applied during traing is given as input and the results are verified.
3.4. Post-processing
Post-processing [4] involves steps to be taken after classification using neural network is
completed. It may include steps like representing the output in Unicode format and also
disambiguation of confusing pairs such as ‘Pa’ and ‘Va’(shown in Figure 5).
Figure 5. Pair of characters in confusion set.
This pair will have almost same direction feature vectors. So some additional disambiguating
technique should be used for such confusing pairs. Eg. In case of ‘Pa’ and ‘Va’, the number of
pixels above and below the horizontal axis can be compared. Such disambiguation technique is
to be devised as post-processing mechanism for every confusing pair identified during the
training of the classifier.
Hence the entire methodology is shown in Figure 6.
Figure 6. Steps involved in training and testing phases of a classifier network

57
4. OUTPUT
Figure 7. Handwritten character being converted to machine readable form
The handwritten Malayalam acquired by a digital pen or stylus will be converted into editable
characters in the computer in one of the recognized fonts of the Malayalam language.
Characters are recognised using Unicode 5.01 or above.
3. CONCLUSION AND FUTURE WORK
The method identified is used to recognize single character at a time. Here each character,
whether it is a consonant character, a vowel character or a dependent vowel symbol (which
usually comes along with consonant character) will be identified as separate character and will
be assigned separate unicodes as shown in Figure 8.
Figure 8. Instead of 0D08, we store it as two separate characters with unicodes 0D07 and 0D57
The handwritten Malayalam acquired by a digital pen or stylus will be converted into editable
characters in the computer in one of the recognized fonts of the Malayalam language.
Characters are recognised using Unicode 5.01 or above.
When trying to extend the system to identify words, an additional step is required during post-
processing which combines these two characters, which actually form a single entity, into a
single character.
Disambiguation of characters can be done based on the position and meaning each character
gives to the word, hence making the system more efficient. Eg. The letters ‘tta’ having Unicode
0D20 and the Malayalam sign anuswara having Unicode 0D02, have the same representation.
They can be disambiguated based on position and neighboring character rules of the language.
Also, additional mechanisms such as automatic completion of the word, spell checker, etc can
be incorporated into the system.

58
REFERENCES
[1] Abdul Rahiman M, M S Rajasree, Masha N, Rema M , Meenakshi R, Manoj Kumar G,
“Recognition of Handwritten Malayalam Characters using Vertical & Horizontal Line Positional
Analyzer Algorithm”, IEEE, pp 268-274, 2011.
[2] Jomy John, Pramod K. V, Kannan Balakrishnan, “Offline Handwritten Malayalam Character
Recognition Based on Chain Code Histogram”, Proceedings ofICETECT, pp 736-741, 2011.
[3] Sreeraj.M, Sumam Mary Idicula,“On-Line Handwritten Character Recognition using Kohonen
Networks”, World Congress on Nature & Biologically Inspired Computing (NaBIC 2009),pp
1425-1430,2009
[4] Suresh Sundaram, A G Ramakrishnan, “An Improved Online Tamil Character Recognition
Engine using Post-Processing Methods”, 10th International Conference on Document Analysis
and Recognition, pp 1216-1220, 2009
[5] Marwan Ali.H. Omer, Shi Long Ma, “Online Arabic Handwriting Character Recognition Using
Matching Algorithm”, IEEE, pp259-262, 2010
[6] Sreeraj.M, Sumam Mary Idicula, “k-NN based On-Line Handwritten Character recognition
system”, First International Conference on Integrated Intelligent Computing, pp 171-176, 2010
[7] http://www.learnartificialneuralnetworks.com/
Authors
Amritha Sampath is a postgraduate engineering student in Computer Science and
Engineering at Rajagiri School of Engineering and Technology, Kochi, India. She completed
her graduation in Computer Science and Engineering and secured high score in Graduate
Aptitude Test in Engineering.
Tripti. C is is working as Assistant Professor in the Department of Computer Science and
Engineering at Rajagiri School of Engineering & Technology, India. She is a postgraduate in
Computer Science and Engineering from CDAC Noida, India and is now pursuing Ph.D
from Cochin University of Science and Technology, in the area of vehicular and Adhoc
networks. She did her graduation in electronics and communication engineering from
Rajagiri School of Engineering and Technology.
Dr. Govindaru V did his Ph.d from ISEC, Bangalore, India. He did his post graduation from
Jawaharlal Nehru University, India. Now he is working as head of Research and
Development Division in C-DIT, Triruvananthapuram, India.

FREEMAN CODE BASED ONLINE HANDWRITTEN CHARACTER RECOGNITION FOR MALAYALAM USING BACKPROPAGATION NEURAL NETWORKS

More Related Content

What's hot (18)

Similar to FREEMAN CODE BASED ONLINE HANDWRITTEN CHARACTER RECOGNITION FOR MALAYALAM USING BACKPROPAGATION NEURAL NETWORKS (20)

More from acijjournal (20)

Recently uploaded (20)

FREEMAN CODE BASED ONLINE HANDWRITTEN CHARACTER RECOGNITION FOR MALAYALAM USING BACKPROPAGATION NEURAL NETWORKS