THE RESULTS OF THE 1ST IPTP CHARACTER-RECOGNITION COMPETITION AND STUDIES ON MULTIEXPERT RECOGNITION FOR HANDWRITTEN NUMERALS

The Institute for Posts and Telecommunications Policy (IPTP) held its first character recognition competition in 1992 to ascertain the prese nt status of ongoing research in character recognition and to find pro mising algorithms for handwritten numerals. In this paper, we report a nd analyze the results of this competition. In the competition, we ado pted 3-digit handwritten postal code images gathered from live mail as recognition objects. Prior to the competition, 2,500 samples (7,500 c haracters) were distributed to the participants as training data. By u sing about 10,000 different samples (29,883 characters), we tested 13 recognition programs submitted by five universities and eight manufact uring companies. According to the four kinds of evaluation criteria: r ecognition accuracy, recognition speed, robustness against degradation , and theoretical originality, we selected the best three recognition algorithms as the Prize of Highest Excellence. Interestingly enough, t he best three recognition algorithms showed considerable diversity in their methodologies and had very few commonly substituted or rejected patterns. We analyzed the causes for these commonly substituted or rej ected patterns and, moreover, examined the human ability to discrimina te between these patterns. Next, by considering the complementary char acteristics of each recognition algorithm, we studied a multi-expert r ecognition strategy using the best three recognition algorithms. Three kinds of combination rules: voting on the first candidate rule, minim al sum of candidate order rule, and minimal sum of dissimilarities rul e were examined, and the latter two rules decreased the substitution r ate to one third of that obtained by one-expert in the competition. Fu rthermore, we proposed a candidate appearance likelihood method which utilizes the conditional probability of each of ten digits given the c andidate combination obtained by each algorithm. From the experiments, this method achieved surprisingly low values of both substitution and rejection rates. By taking account of its learning ability, the candi date appearance likelihood method is considered one of the most promis ing multi-expert systems.