Symbolic Segmentation of Handwritten Numerals with Robust Fuzzy Clustering

This paper presents a new approach to symbolic handwritten numerals segmentation and pre-processing. It is based on recent developpments i n robust fuzzy clustering. Adaptative linear shape detection is achieved taking into account noise caracterisation . Introduction This work is a part of more complete project of handwritten numerical amount of cheque interpretation. We will present work related to detection of linear shapes considered as fuzzy clusters. The basic fuzzy clustering algorithm, FCM (Fuzzy C-Mean) [ I ] generates a fuzzy partition providing a measure of the membership degree of each pixel or "pattern" to a given region or cluster. Most applications in computer vision were in image segmentation [ 2 ] . Fuzzy models allows efficient contextual decision [3, 41. interested by specific shape detection, this algorithm becomes ineffective. Recent development extend this algorithm to the case of linear or shell-like clusters performing segmentation and fitting simultaneously . It was proved that perfomance of these algorithms were more interesting than those of Hough transform. Furthermore, fuzzy modeling provides very rich and useful information, in our approach, of membership degrees. Adaptative linear shape detection, takes into account the lenght and the extent of each linear cluster. We introduce a rejection linear cluster that collects noise and outliers. This idea of noise cluster was firstly introduced by [ 5 ] . Interpretation of handwritten numerical amount of cheque require some pre-processing, essentielly removing basic line of cheque on which we wrote the amount.This method allows us to detect and then to remove this line without damage to numerals. Membership degrees information allows efficient separation between numerals and this line providing efficient symbolic representation of numerals. Fig1 shows symbolic segmentation of numerals that are compound of linear features. Number of clusters is not crucial since we can overestimate and achieve cluster merging according to fusion criterion Fig2. Fig3 shows ambiguous pixels (in light grey) that belongs to numerals and basic cheque line Handwritten numerals simultaneously in a synthetic case. segmentation and linear shape This information of membership detection degree allows us to remove line In its basic version, FCM generates without truncate numerals. Fig 4 spherical filled clusters. If we are shows result on real amount. E-mail : Nozha.Boujemaa(ujinria.fr Conclusion + 64 Avenue Jean Portalis, 37200 Tours France This paper presents an original and 19 Rue de la vallee Maillard B.P. 13 1 1 a promising way to handwritten 4 10 13 Blois Cedex France numerical amount of cheque interpretation. Shared pixels gives a very useful information to hand le uncertainty. More results a re i n progress. [ I ] J . C BEZDEK, "Pattern recognition with fuzzy objective function algorithms" (Plenum PressNew York 198 1) [2] J .C. BEZDEK & S. PAL, "Fuzzy Models for Pattern Recognition : methods tha t search for s t ruc tures in data" (IEEE P r e s s 1992). Application to Medical Image Segmentation" in Progress in Image Analysis and Processing 111, pp: 649656. Ed. S. Impedovo World Scientific 1994. [4] N. BOUJEMAA, G. STAMON, J. LEMOINE and E. PETIT, "Fuzzy ventricular Endocardium Detection with Gradual focusing decision", International Conference of the IEEE Engineering in Medecine and Biology Societyvol. 14, pp. 1893-1894, Paris Octobre 92. [5] R. DAVE, "Characterisation and detection of noise in clustering", Pattern Recognition Letters, vo1.12, pp. 657-664, NOV 1991. [3] N. BOUJEMAA and G. STAMON, "Fuzzy Modeling in Early Vision Fig. 1 linear segmentation of numerals Fig.2 Cluster merging result (Ambigous pixels are in light grey).