Recognition of Words from Legal Amounts of Indian Bank Cheques

Legal amount of Indian bank cheques contains 36 different words. Most of the Indian cheques in cities are written in English although some of them are written in Hindi and other state languages. As the legal amount words written in English can be case sensitive, the size of the lexicon for legal word recognition can go up to 108 (3×36). In this paper a lexicon driven segmentation-recognition scheme is proposed for the recognition of legal amount words from Indian bank cheques written in English. A water reservoir concept is used to pre-segment the words into primitive components and the primitive components of a word are then merged into possible characters to get the best word using the lexicon of 36 different legal words of bank cheque. To merge these primitive components into characters and to get optimum character segmentation, dynamic programming is employed using total likelihood of the characters of a word as an objective function. To calculate the likelihood of a character, Modified Quadratic Discriminant Function (MQDF) is used. The features used in the MQDF are mainly based on directional features of the contour points of the components. In the paper it is assumed that the words are already extracted from the cheque image for recognition. A database consisting of 5400 words, collected from 50 writers has been used for testing the system and an accuracy of 97.04% was observed.