Minimally Segmenting High Performance Bangla Optical Character Recognition Using Kohonen Network

This paper presents a method to use Kohonen neural network based classifier in Bangla Optical Character Recognition (OCR) system, providing much higher performance than the traditional neural network based ones. It describes how Bangla characters are processed, trained and then recognized with the use of a Kohonen network. While there have been significant efforts in using the various types of Artificial Neural Networks (ANN) in optical character recognition, this is the first published account of using a segmentation-free optical character recognition system for Bangla using a Kohonen network. The methodology presented here assumes that the OCR pre-processor has minimally segmented the input words into easily segmentable chunks, and presenting each of these as images to the classification engine described here. The size and the font face used to render the characters are also significant in both training and classification. The images are first converted into grayscale and then to binary images; these images are then scaled to a fit a pre-determined area with a fixed but significant number of pixels. The feature vectors are then extracted from the rectangular pixel map, which in this case is simply a series of 0s and 1s of fixed length. Finally, a Kohonen neural network is chosen for the training and classification process. Although the steps are simple, and the simplest network is chosen for the training and recognition process, the resulting classifier is accurate to better than 98%, depending on the quality of the input images. 1