Multi-font English Character Recognition based on Modified Invariant Moments

Abstract

This paper describes an approach based on modified invariant moments for recognition of multi-font English characters. The proposed method is independent of size and translation variations and shows better results under noisy conditions. The work treats isolated English characters which are normalized to a size of $33\times 33$ pixels and the image is thinned. As a pre-classification step, end points and Euler numbers have been estimated from this thinned image of the character. For size and translation invariance, the modified invariant moments suggested by Palaniappan have been evaluated. The system is trained for 7 different font styles with 364 images. A decision tree-based minimum distance nearest neighbor classifier has been adopted for classification. The system is tested for these seven fonts with various sizes of the characters between 8 to 72. A total of 7,280 character images are tested with this system and the success rate is found to be 99.65\%. The method shows encouraging results on multi-font/sized character images.