TY  - JOUR
T1  - Structure Analysis of Multilayer Perceptron Network for Handwritten Tamil Character Recognition Using Levenberg-Marquardt Algorithm
AU - , J. Sutha AU - , N. Ramaraj 
JO  - International Journal of Soft Computing
VL  - 3
IS  - 5
SP  - 373
EP  - 381
PY  - 2008
DA  - 2001/08/19
SN  - 1816-9503
DO  - ijscomp.2008.373.381
UR  - https://makhillpublications.co/view-article.php?doi=ijscomp.2008.373.381
KW  - Handwritten character recognition
KW  -preprocessing
KW  -segmentation
KW  -feature extraction
KW  -fourier descriptors
KW  -transition values
KW  -average pixel values
KW  -distances and angle features
KW  -LM algorithm
AB  - This study addresses the problem of recognizing handwritten Tamil characters, the popular south Indian Language. Applying neural network for recognizing a large volume of high dimensional data is a difficult task as the training process is computational expensive. Back propagation(BPN) training algorithm has been applied to various problems by researchers. However, the main drawback in applying that algorithm for real world problem is the low convergence speed in the entire learning process. This proposed system uses Levenberg-Marquardt (LM) algorithm for training a multilayer perceptron (MLP) network with one hidden layer. It is a Hebbian-based algorithm then the training process converges quickly compared to BPN algorithm. Various preprocessing algorithms and feature extraction techniques performed prior to the recognition of handwritten Tamil characters are also implemented. The performance of the LM algorithm has been checked for the recognition of handwritten Tamil characters and structure analysis is carried out to find the optimum structure of the MLP network. Test results indicate that the proposed system provide good recognition accuracy of 94.1% for handwritten Tamil characters and structure analysis suggest that the appropriate structure should have 50 hidden nodes with tangent sigmoid function for hidden nodes, pure linear function for output nodes with a maximum training epoch of 150 to achieve the higher recognition rate for the MLP network trained using LM algorithm.
ER  -