Automatic reading of the literal amount of bank checks

Abstract

The studies presented in this paper deal with the global recognition of a restricted variety of handwritten words comprising the vocabulary used to write French bank checks. Several tools have been developed within the constraints of this application, tools that relate to the general problem of off-line cursive handwriting recognition. The first difficulty when one wants to read a text is the location of words. This is done by using a top-down analysis that first locates lines of text before segmenting them into individual words. A structural model of cursive handwriting, built on the median axis of the word, is proposed. This constitutes an alternative to the use of the conventional, letter-by-letter analytical model and could be used for other problems involving cursive word recognition. According to this structural model of cursive handwriting, an edit distance is computed between the extracted structural description and the reference descriptions that are interpreted as grapheme strings. This provides an ordered list of candidates for each individual word. Feature extraction in the binary image of the word is performed using a specific line-following algorithm. Since it is possible to express the syntax of the sentences by a finite grammar, this information is used to discard the inconsistent sentences from the possible ones. These various algorithms have been tested on personal data, as well as on real check images.