Cursive script recognition by backward matching

This paper proposes a new model for cursive script recognition which is both analytical and global and emphasizes the role of high-level contextual information. This model is based both on a top-down recognition scheme called Backward Matching and a bottom-up feature extraction process which are working in a competitive way. The top-down recognition scheme allows multi-level correspondence between the levels of representation of the image word and those of the symbolical descriptions of the lexicon. (This work was initiated at the IBM Almaden Research Center, CA, USA)