A Neural Approach to OCR System Design

Handwriting is a natural way to communicate and record information. Machine Simulation to recognize off-line handwriting has opened new horizons to improve human-computer interface. This book presents superior approaches for character image pre-processing, untouched as well as touched character segmentation and feature extraction for the purpose of handwritten word recognition experiment. The first segmentation technique is based on the connected component analysis and is proposed to segment untouched characters in a word image and in the second technique, a heuristic vertical dissection based approach is proposed to segment touched characters in a word image. A fusion of two feature extraction techniques i.e Binarization and Projection Profile Techniques is used to evaluate the performance of the two variants of Artificial Neural Networks, namely, Feed Forward Back Propagation NN and Radial Basis Function NN in terms of accuracy, speed and computational complexity. To help the researchers, various techniques to optimize the training parameters of an ANN are evaluated and some common situations during BP Learning, with possible causes and potential remedies are also presented.