Recognition of Multi-font English Numerals using SOM Neural Network

Abstract

In this paper a new scheme is proposed for off-line recognition of multi-font numeral, using neural networks. Recognition of numerals has been a research area for many years because of its various applications. But there wasn’t much research done for recognition of multi-font numerals. The approaches proposed so far, suffer from larger computation time and training because they must have a set of training samples per each font. They can be extended to recognize many more fonts but the accuracy decreases rapidly. So as to eliminate these drawbacks, in this paper, a method is presented which recognizes 30 different fonts of different sizes varying from size 10 to 28, with an accuracy of 99.55% on a database of 2000 numeral images. The purpose of this study is to provide a new method to recognize digits based on neural network that can identify the same symbols after training without limitation on the type of the font. In the proposed method, a high accuracy rate is achieved in recognizing digits by extracting the appropriate features without the need for complex neural network architecture. This method uses a self-organizing map (SOM) neural network to measure similarity between the features of digits and the features of the indicators associated with the digits from 0 to 9 obtained in the training stage. In this method, one sample is used for each digit to train the network. So, the proposed method can be used to recognize typed letters without limitation on fonts. General Terms Pattern Recognition, Neural Networks, OCR