Fuzzy-Shape Grammars for Cursive Script Recognition

Abstract

This paper presents a new approach for cursive script segmentation and recognition, based on intrinsic models of cursive letters (allographs). The models are built using strati ed context-free shape grammars that permit the definition of both syntactic and semantic attributes. These attributes synthetize pertinent morphological characteristics of allographs that are then used for recognition. The main topic of this paper concerns the parsing process developed for allograph segmentation, which uses fuzzy-logic to evaluate the likelihood of segmentation hypotheses. This process is the rst step of the recognition method and leeds to the construction of a graph where nodes represent segmented allographs and arcs link adjacent nodes. The analysis of this segmentation graph can be carried out for submitting possible letter sequences to higher linguistic evaluation modules. Preliminary results are given for multi-writer isolated cursive letters. For a test database containing cursive samples of 10 di erent writers, an average recognition rate of 91:7% is obtained. Recognition is non personalyzed, that is, cursive samples of all writers are treated with the same algorithm parameters. This work was published in Advances in Structural and Syntactic Pattern Recognition, H. Bunke (editor), World Scienti c Publishing, pp. 320-332, 1993. It was supported in part by NSERC Canada under grant OGP000915, and in part by FCAR Qu ebec under grant CRP2667. Marc Parizeau received scholarships from NSERC and FCAR. Marc Parizeau was with the D epartement de g enie electrique et de g enie informatique, Ecole Polytechnique de Montr eal. He is now with the D epartement de g enie electrique et de g enie informatique, Universit e Laval, St-Foy (PQ), Canada, G1K 7P4. E-mail: parizeau@gel.ulaval.ca R ejean Plamondon is with the D epartement de g enie electrique et de g enie informatique, Ecole Polytechnique de Montr eal, C.P. 6079 { Succ. \A, Montr eal, Canada, H3C 3A7. Guy Lorette is with the Institut de Recherche en Informatique et Syst emes Al eatoires, Universit e de Rennes I, 35042 Rennes Cedex, France