Word - level Deciphering Algorithm for egraded Document Recognition

A text recognition algorithm is proposed that uses word-level language constraints in a deciphering framework to directly decode the identity of each word pattern in an input text. This is a font-independent approach that solves problems of touching characters and character fragmentation. The major difficulty of using a deciphering approach on the word level is that the existence of relatively stable and reliable language constraints on the character level, such as character n-grams and a vocabulary of common words, usually do not scale up to the word level. A word-level deciphering approach is presented in this paper that solves a selected portion of an input text using a word-level relaxation deciphering algorithm. Font information is then learned from the solved portion of the text and used to re-recognize the rest of the text. Tests of the proposed approach on both artificially generated and scanned documents show satisfactory performance in the presence of image degradation.