1. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
INTERNATIONAL JOURNAL OF ELECTRONICS AND
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
COMMUNICATION ENGINEERING & TECHNOLOGY (IJECET)
ISSN 0976 – 6464(Print)
ISSN 0976 – 6472(Online)
Volume 4, Issue 6, November - December, 2013, pp. 117-123
© IAEME: www.iaeme.com/ijecet.asp
Journal Impact Factor (2013): 5.8896 (Calculated by GISI)
www.jifactor.com
IJECET
©IAEME
EVALUATING NEURAL NETWORK AND HIDDEN MARKOV MODEL
CLASSIFIERS FOR HANDWRITTEN WORD RECOGNITION
Chaitali K.Dhande1, P.M.Mahajan2
1
2
Dept of E&Tc, J.T.Mahajan C.O.E.North Maharasthra Univresity, Faizpur India
Dept of E&Tc, J.T.Mahajan C.O.E.North Maharasthra Univresity, Faizpur India
ABSTRACT
One of the most classical applications of the Neural Network and Hidden Markov Model is
the Handwritten Word recognition. This system is base for many different types of applications in
various fields, many of which we use in our daily lives. The proposed system would include a simple
scheme for two different classifiers .An Neural Networks (NN) classifier based on Multi-Layer
Perceptron (MLP) trained with Back Propagation algorithm and Hidden Markov Models (HMM)
classifier with two states. HMM classifier is used to identify character in sequence of word with
assigning a probability to each of them. NN classifier is used to generate a score for each segmented
character. In the end the score from NN and HMM classifier will be compared to get optimized
performance. To take the advantages of good properties of both methods, proposed approach will
compare the result obtained by NN and HMM
Keywords: HMM, NN, MLP, Back Propagation algorithm
1. INTRODUCTION
Handwriting word recognition is the ability of a computer to receive and interpret
handwritten input from sources such as paper documents, photographs, touch-screens and other
devices. According to the way of sensing, Handwriting Word Recognition can be categorizing as Online recognition and Off-line recognition. Off-line handwriting recognition involves the automatic
conversion of text in an image into letter codes which are usable within computer and textprocessing applications. In the proposed system off line handwritten word recognition is adapted
[3,13].
1.1 Problem Definition
Word recognition can be defined as the categorization of the input character images into
identifiable classes via extraction of significant features or attributes of the character from the
117

2. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
background of irrelevant details. The word recognition system is the process of identifying the
sequence of alphabets .The sequence (i.e. word) must be present in database. System should identify
all probable combinations of available images and to identify most appropriate word from dictionary.
The system refers to two different classifiers:
1. Hidden Markov Model (HMM) Classifier
2. Neural Network (NN) Classifier
The system includes a simple scheme for the integration of two different classifiers. NN
classifier based on Multilayer Perceptron (MLP) trained with Back Propagation Algorithm and the
HMM classifier with two states.HMM classifier is used to identify characters in sequence of word
with assigning a probability to each of them. An NN classifier is used to generate a score for each
segmented character. In the end the scores from HMM and NN classifier are compared to get
optimized performance. An approach is proposed to compare results of NN and HMM, that takes the
advantages of good properties of both methods. Finally, the most appropriate word is reordered
according to the new composite scores.
1.2 Current State of Arts
Handwritten word recognition techniques can be roughly classified into two different approaches:
● Segmentation based approaches: segmentation based approaches have been used widely both
for off-line and on-line handwriting recognition.
● Segmentation free approaches: Segmentation free approaches bypass the segmentation phase.
In Segmentation based approaches, the word is segmented in various parts and then further
processing takes place. Every segment is acting as the one of the element of sequence. Every
segmented element can be called as the sub-word unit. Word is ordered sequence of such units.
Many systems use HMMs to model sub–word units (characters) and the Viterbi algorithm to find the
best match between a sequence of observations and the models [5, 6, 9]. Segmentation based
approach is more used in the systems like spelling correction or postal address recognition etc.
During the last few years, HMMs have become a very popular approach in handwriting recognition.
One of the reasons is their higher performance in medium to large vocabulary applications where
segmentation–recognition methods are used. Such methods cope with the difficulties of segmenting
words into characters.
The Viterbi algorithm is optimal in the sense of maximum likelihood and it looks at the
match of the whole sequence of features (observations) before deciding on the most likely state
sequence. This is particularly valuable in applications such as handwritten word recognition where
an intermediate character may be garbled or lost, but the overall sense of the word may be detectable.
On the other hand, the local information is somewhat overlooked. Furthermore, the conditional–
independence imposed by the Markov Model (each observation is independent of its neighbors)
prevents an HMM from taking full advantage of the correlation that exists among the observations of
a single character [10].
2. RELATED WORK
Lawrence [7] describes review of theoretical aspect of HMM as a statistical modeling and
show how they are applied to selected problems in machine recognition of speech. In this paper he
attempted to present the theory of hidden Markov models from the simple concept of discrete
Markov chain. He also attempted to illustrate some applications of the theory of HMMs to simple
118

3. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
problems in speech recognition and pointed out how the techniques have been applied to more
advanced speech recognition problem.
Tay etal [12] describes an approach to combine neural network (NN) and Hidden Markov
models (HMM) for solving handwritten word recognition problem. To recognize a word, the NN
computes the observation probabilities for each letter hypothesis in the segmentation graph. The
HMMs then compute the likelihood for each word in the lexicon by summing the probabilities over
all possible paths through the graph. They introduce the discriminant training to train the NN to
recognize junk. They use three database namely IRONOFF, SRTP and AWS. Also show the
superiority of the hybrid recognizer compared to out baseline recognizer, which is using discrete
HMM. Finally, they show that the hybrid recognizer can be bootstrapped automatically from the
discrete HMM recognizer, and significantly improve its recognition accuracy by going through
several training stages.
Koerich etal [8,9] they describes a hybrid recognition system that integrates hidden Markov
models (HMM) with neural networks (NN) in a probabilistic framework. An NN classifier is used to
generate a score for each segmented character and in the end, the scores from the HMM and the NN
classifiers are combined to optimize performance. Experimental results show that for an 80,000–
word vocabulary, the hybrid HMM/NN system improves by about 10% the word recognition rate
over the HMM system alone.
3. METHODOLOGY
The proposed system is Off-line handwriting recognition system which involves the
automatic conversion of an image of text into letter codes which are usable within computer and textprocessing applications.
3.1 System Description
The system works in two phases, Training phase and Execution phase. In Training phase
both the classifiers gets trained with data available and in Execution phase system takes input
preprocess it and both classifiers gets executed to recognize the word.
3.2. Training to Neural Network (Multi Layered Perceptron) Classifier
The Neural Network is trained by Error Back Propagation Training (EBPT) algorithm in the
system. It trains three layers of MLP i.e. input layer, hidden layer and output layer. Note that, first
error signal vector is determined at output layer and then it is propagated towards network input
nodes. The Error Back Propagation Training algorithm is working as follows [1].
Step 1] Initial weight matrix W and V are initialized at small random values.
Step 2] Training starts here. Input is presented and the layers output computed.
Step 3] Error value for given input is computed.
Step 4] Error signal vector for both , output and hidden, layer is computed
Step 5] Output layer weights are adjusted accordingly.
Step 6] Hidden Layer weights are adjusted.
Step 7] step 2 to step 6 executes predetermined ‘n’ times till maximum error is greater than error
value computed in step 3.
Step 8] Training Cycle is completed
3.3. Training to Hidden Markov Models (HMM) Classifier
HMM plays important role in the system. Many systems uses HMMs to model sub-word
units and the Viterbi Algorithm [7,8] to find the best match between the sequence of observation &
the models. As HMM is the static model it must needed to frame with finite states. For a word
119

4. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
recognition system all symbols are considered as states and each word is a particular sequence of the
states. For English language there are 26 symbols and every word is combination of some of this
symbol. Words may contain one or more alphabets It becomes very rigid in HMM to calculate Initial
State Probabilities with more symbols and states. Hence for convenience, in this report, three
symbols (A, B, C) and only three letter words are considered.
The detailed Block Diagram of Execution phase as shown in fig.
Input (Word)
Preprocessing
Segmentation
Feature Extraction
MLP classifier
Character
HMM classifier
set
Character Sequence
probability
Combination of
results and
recognition
Fig. 1 Detailed block diagram of system
Preprocessing: The input of preprocessing is the image which is a scanned handwritten word.
Preprocessing makes the word ready for further processing. Noise Reduction in image processing it
is usually necessary to perform a high degree of noise reduction in an image before performing
higher-level processing steps. In the presented work removal of noise helps in segmentation and
feature extraction. In the system the Gaussian filter is used.
Resizing: As for segmentation, in system histogram based algorithm is used, it demands fix size of
input image. Hence here the resizing of input image is necessary to bring image in predetermined fix
size. For resizing the median is calculated and from it the certain distance is captured.
120

5. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
Segmentation: In the system, Histogram-based method is used. In the system the all pixel values are
summed column wise and depend on the sum values the segmentation is done. The system, segment
the word into multiple regions and considers a region at a time to provide as input to HMM
classifier.
Feature Extraction: When the input data to an algorithm is too large to be processed and it is
suspected to be notoriously redundant then the input data will be transformed into a reduced
representation set of features In the system every segmented image i.e. a character region pixel
values are considered as feature vector for that segment. For all segments such feature vectors can be
drawn. All feature vectors are of same size.
NN Classifier: Extracted features of individual character are provided to the NN. It provides
recognition of each character and provides the set of characters at respective output nodes.
HMM Classifier: It also works on the same features extracted of input image. It identify characters
by feature and then checks the probability of sequence. If not found in history, HMM suggests
nearest sequence possible.
4. RESULTS
The system is implemented using MATLAB 7.0
Recognition of word is as follows
System provides two end results
1. No.of words guessed by system which has maximum probability w.r.t previous data
provided
2. The nearest combination of characters
121

6. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
No. of Input Word
Recognition by HMM
Recognition by NN
26
51
77
103
25
49
75
100
20
45
70
90
The results obtained shows comparision of both the methods by taking no.of input words.
Efficiency of HMM is better than NN. The system results vary if segmentation phase is not giving
proper results.
5. CONCLUSION
From Results, we can conclude that the system can be proved as a efficient handwritten word
recognition system. An NN classifier is trained to identify each segmented character, so the
interpersonal feature variations do not affect the word recognition results. Whereas the HMM
classifier is trained for checking the sequence of characters, so wrong sequences can also guess to the
proper word. If the HMM is trained for any symbols in any language, it can recognize the proper
sequence. But there are some limitations also, main is the rigid structure of HMM and its dependency
on the history data. There may be some perfect answer which cannot be guessed by HMM, just
because it is not in history or rarely applied in history. NN classifier which do not guess at all. It just
generates the answer for what it is trained for. The error limit in EBPT (Error Back Propagation
Training) algorithm may vary some results.
Both HMM and NN are having very high dependency on the features which used to train
them and because of that individual performance of both systems fails in the case of large
vocabulary. Combination of both classifiers gives more accurate results even for large vocabulary.
Further Works
The system can be used in various applications, and for many languages. Only the need is to
train the HMM's for those symbols. We can obtain better results if size of history data is increased.
The system can be further utilizing in optical word recognition, huge handwritten data processing
centers etc. Not only the Word Recognition system, but same HMM logic can be used for the Speech
Recognition or pattern identification also.
REFERENCES
Books
[1]
[2]
Jackel M. Zurada. Introduction to Artificial Neural Systems, Jaico Publishing House, 2004,
pg.185-pg.235.
Rafael C. Gonzalez and Richard E Woods. Digital Image Processing, second edition, Pearson
Education, 2005, pg.589-pg.658.
Theses
[3]
A. Senior. Off–Line Cursive Handwriting Recognition using Recurrent Neural Networks.
PhD thesis, University of Cambridge, Cambridge, England, September 1994.
122

7. International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 –
6464(Print), ISSN 0976 – 6472(Online) Volume 4, Issue 6, November - December (2013), © IAEME
Proceeding Papers
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
Fu Chang, Chin-Chin Lin and Chun-Jen Chen, Applying a hybrid method to handwritten
character recognition Institute of Information Science, Academia Sinica, Taipei, Taiwan
M. Y. Chen, A. Kundu, and S. N. Srihari. Variable duration hidden markov model and
morphological segmentation for handwritten word recognition. IEEE Transactions on Image
Processing, 4(12):1675–1688, 1995.
A. El-Yacoubi, M. Gilloux, R. Sabourin, and C. Y. Suen. Unconstrained handwritten word
recognition using hidden markov models. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 21(8):752–760, 1999.
Lawrence R. Rabiner, A Tutorial on Hidden Markov Model and Selected Applications in
Speech Recognition, Proceedings of the IEEE, vol. 77, no.2, February 1989.
A. L. Koerich, R. Sabourin, and C. Y. Suen. Fast two–level viterbi search algorithm for
unconstrained handwriting recognition. In Proc. 27th International Conference on Acoustics,
Speech, and Signal Processing, Orlando, USA, 2002. To appear.
A. L. Koerich, R. Sabourin, C. Y. Suen, and A. El-Yacoubi. A syntax–directed level building
algorithm for large vocabulary handwritten word recognition. In Proc. 4th International
Workshop on Document Analysis Systems, pages 255–266, Rio de Janeiro, Brasil, 2000.
G. Zavaliagkos, Y. Zhao, R. Schwartz, and J. Makhoul. A hybrid segmental neural net/hidden
markov model system for continuous speech recognition. IEEE Transactions on Speechand
Audio Processing, (1):151–160, 1994.
John A. Fitzgerald, Bing Quan Huang, and Tahar Kechadi. An Efficient Hybrid Approach
for Online Recognition of Handwritten Symbols. Department of Computer Science
University College Dublin Belfield, Dublin 4, Ireland.
Yong Haw Tay’, Pierre-Michel hllicad, Marzuki Khalid’, Christian Viard-Gaudin3 Stefan
Knerr, Offline Handwritten Word Recognition Using A Hybrid Neural Network And Hidden
Markov Model, International Symposium on Signal Processing and its Applications (ISSPA),
Kuala Lumpur, Malaysia, 13 - 16 August, 2001. Organized by the Dept. of Microelectronics
and Computer Engineering, UTM, Malaysia and Signal Processing Research Centre, QVT.
Australia,
Yong Haur Tay’, Pierre-Michel Lallican*, Marzuki Khalid’, Stefan Knerr2, Christian WardGaudin3 , An Analytical Handwritten Word Recognition System with Word-level
Discriminant Training, 0-7695-1263-1/011$10.00 0 2001 IEEE
Yong Haw Tay', Pierre-Michel Lallican2, Marzuki Khalid', Christian Viard-Gaudin3, Stefan
Kneer, An Offline Cursive Handwritten Word Recognition System , IEEE Catalogue No. 01
CH37239 0-7803-7101-I/01/$10.000 2 001 IEEE.
Gunjan Singh, Avinash Pokhriyal and Sushma Lehri, “Fuzzy Rule Based Classification And
Recognition Of Handwritten Hindi Curve Script”, International Journal of Computer
Engineering & Technology (IJCET), Volume 4, Issue 1, 2013, pp. 337 - 357, ISSN Print:
0976 – 6367, ISSN Online: 0976 – 6375.
Primekumar K.P and Sumam Mary Idicula, “Performance of On-Line Malayalam
Handwritten character Recognition using HMM and SFAM”, International Journal of
Computer Engineering & Technology (IJCET), Volume 3, Issue 1, 2012, pp. 115 - 125,
ISSN Print: 0976 – 6367, ISSN Online: 0976 – 6375.
123