Recognition of handwritten numerals using topological structures and Fourier descriptors of numeral contours

A new recognition structure for handwritten numerals is proposed. The method is basically a multistage structure combined with a parallel recognition classifier, which is based on the topological properties of the characters, statistical analysis, the positions of centroids of outer and interior contours, and the distance of the furthest point of the character outer contour from the centroid. This new scheme eventually classifies input characters into one of 14 subgroups. This recognition structure is more effective than a single-stage classifier, especially for handwritten characters, because of the vast variations in writing styles. After preprocessing (broken-character mending, gap filling, noise removal and contour smoothing) and contour following, a handwritten character is classified into one of three groups according to its topological properties. Then, each group is further classified into subgroups and so on. After classification as to which of the 14 subgroups the input character belongs to, the final recognition is based on a comparison of the Fourier descriptors of the outer contour of the input character and those of prototypes of that subgroup. In our experiment, 1000 characters from the NIST database were used for the statistical analysis, and 427 prototypes representing different writing styles were chosen for comparison; then 5841 different characters were used for testing. The recognition rate obtained was 98.31%, with a reliability of 99.12%, a substitution rate of 0.87% and a rejection rate of 0.82%.