Through comparing the results of two successive IPTP Character Recognition Competitions which focused on 3-digit handprinted postal codes, we herein analyze the methodologies of the submitted algorithms along with the substituted or rejected patterns of these algorithms. Regarding their methodologies, lesser diversity was apparent specifically concerning the contour-chain code based on local stroke directions and statistical discriminant functions for feature extraction and discrimination. Analysis of the patterns demonstrated that the misrecognized patterns being most often improved were categorized as a decrease in peculiarly shaped handwritten characters or heavy-handed and disconnected strokes. However, most of the remaining misrecognitions were still classed as peculiarly shaped handwriting as commonly shared between the best three algorithms. From these analyses, we could delineate a direction to be taken for developing more effective methodologies and clarify the remaining problems to be overcome by the subsequent intensive research. Furthermore, we evaluate in this article our multi-expert recognition system for achieving higher recognition performances by means of combining complementary recognition algorithms. We performed a subsequent investigation of the Candidate Appearance Likelihood Method using novel experimental conditions and a new examination of the application of the neural network as the combining method for accumulating the broader candidate appearances. The results obtained confirm that combining through the neural network constitutes one of the most effective ways of making the multi-expert recognition system a reality.

This paper describes an on-line Kanji character recognition method that solves the one-to-one stroke correspondence problem with both the stroke-number and stroke-order variations common in cursive Japanese handwriting. We propose two kinds of complementary algorithms: one dissolves excessive mapping and the other dissolves deficient mapping. Their joint use realizes stable optimal stroke correspondence without combinatorial explosion. Also, three kinds of inter-stroke distances are devised to deal with stroke concatenation or splitting and heavy shape distortion. These new ideas greatly improve the stroke matching ability of the selective stroke linkage method reported earlier by the authors. In experiments, only a single reference pattern for each of 2,980 Kanji character categories is generated by using training data composed of 120 patterns written carefully with the correct stroke-number and stroke-order. Recognition tests are made using the training data and two kinds of test data in the square style and in the cursive style written by 36 different people; recognition rates of 99.5%, 97.6%, and 94.1% are obtained, respectively. Moreover, comparative results obtained by the current OCR technique as applied to bitmap patterns of on-line character data are presented. Finally, future work for enhancing the stroke matching approach to cursive Kanji character recognition is discussed.

The Institute for Posts and Telecommunications Policy (IPTP) held its first character recognition competition in 1992 to ascertain the present status of ongoing research in character recognition and to find promising algorithms for handwritten numerals. In this paper, we report and analyze the results of this competition. In the competition, we adopted 3-digit handwritten postal code images gathered from live mail as recognition objects. Prior to the competition, 2,500 samples (7,500 characters) were distributed to the participants as traning data. By using about 10,000 different samples (29,883 characters), we tested 13 recognition programs submitted by five universities and eight manufacturing companies. According to the four kinds of evaluation criteria: recognition accuracy, recognition speed, robustness against degradation, and theroretical originality, we selected the best three recognition algorithms as the Prize of Highest Excellence. Interestingly enough, the 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 rejected patterns and, moreover, examined the human ability to discriminate between these patterns. Next, by considering the complementary characteristics of each recognition algorithm, we studied a multi-expert recognition strategy using the best three recognition algorithms. Three kinds of combination rules: voting on the first candidate rule, minimal sum of candidate order rule, and minimal sum of dissimilarities rule were examined, and the latter two rules decreased the substitution rate to one third of that obtained by one-expert in the competition. Furthermore, we proposed a candidate appearance likelihood method which utilizes the conditional probability of each of ten digits given the candidate 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 candidate appearance likelihood method is considered one of the most promising multi-expert systems.