Recognition Of Cursive Texts Using Hamming Neural Nets

Abstract

Hamming Neural networks are employed for cursive text character separation and recognition. The digitized image of the scanned text is first enhanced by applying simple contrast stretching algorithm. Image registration is then performed to eliminate the white margins. Lines are separated by rows of white pixels and the morphological components identified. The projection method in conjunction with histogram analysis is used to estimate the width of each character in a word. The word is approximately decomposed into its constituent characters. The Hamming net is used to identify thus separated characters, while supposing the included portions of the adjacent characters as noise. The property of the trained Hamming net, i.e., the associative feature for detecting the best matched patterns from the prototype in a mean square sense, is the basis for constructing the recognition algorithm. Illustrative examples are presented and it is shown that the proposed approach is fast and efficient to provide an on-line separation and classification system for cursive texts such as Persian (Farsi), Arabic, Urdu and English script.