Abstract

Recognition of off-line isolated handwritten English block letters and numerals was done within the constrained domains separately. Distance measurements such as Nearest Neighbor (2), Similarity, Hamming, Linear Correlation, Cross Correlation, Entropy and Information Content were used for recognizing handwritten letters and numerals. 65 prototypes of English block letters and 31 prototypes of numerals were used. Experiments were conducted on over 180 images of letters and 120 images of numerals. In the case of letters it was observed that the results of Nearest Neighbor distance measurement were outstanding and Nearest Neighbor outperforms all other distance measurements. In the case of numerals no distance measurement appeared to be sufficient. In both cases (letters and numerals) Entropy and Information Content does not give any correct results. The thinning algorithm by Zhang & Suen was performed on isolated characters for further improvements. It was observed that thinning with distance measurements has no effect on improving the recognition results. Since no distance measurement appeared to be sufficient in the case of numerals, a neural network of type Multi Layer Perceptron (MLP) was used to combine the results from 6 distance measurements (Nearest Neighbor (2), Similarity, Hamming, Linear Correlation, Cross Correlation) to obtain a single recognition result. It was observed that the combination enhanced the success rate. Further improvements to MLP results were obtained by using a structural verifier