There are two purposes in this paper. The one is to compare experimentally the two types of methods of writer identification, structure analysis and pattern matching, and the other is to examine the dependency of identifiability on letters. The material was provided by seven persons who handwrote a sentence of twenty-six letters twenty-five times repeatedly. About a half of the material was used as a sample for learning and the rest was for test. Templates representing persons were constructed from the sample for learning through each of five methods, three of structure analysis S1, S2, S3 and the other two of pattern matching P1, P2. For each character in the sample for test, the distances from it to the templates were evaluated and the person who was the nearest was judged as the writer. Experimentally the average correct identification rate for P1, which is better than P2, was about 85%, while that for S2, which is better than S1 and S3, was about 80%. As far as the methods of structure analysis proposed by us are concerned, they are inferior to the methods of pattern matching. As for sentences, all of the writers were correctly identified. The correct identification rates for HIRAGANA letters are in general greater than those for KATAKANA, which seems to imply that complicated letters transmit us more information about writers than simple letters do.

This paper proposes an off-line text-independent writer identification method applicable to Japanese and Korean sentences. It is assumed that the writer of a writing in question exists in a certain group of people and that reference writings written by each person in the group can be used for identification. In the proposed method, relative frequencies of some model patterns are counted on the binary pattern of each writing and are used as the feature to measure the distance between two writings. Based on a modified Mahalanobis' distance for this feature, the person whose reference writing is nearest to the writing in question is judged as the writer. The effectiveness of the proposed method is examined through an experiment using Japanese and Korean writings. Error rates in the experiment were different depending on conditions such as volume of reference writings, dimension of adopted features, and number of people to be identified. In some cases, error rates as low as 0% were observed. Error rates tend to be lower in Korean writings probably because Hangul is composed of a smaller number of letters compared to Kanji and Hiragana in Japanese writing.

This paper deals with an on-line signature verification system. It is assumed taht the system requires a person approaching it to declare his name and to write his signature. The system compares the written signature with reference signatures registered in advance and admits his access if the dissimilarity is below a threshold. New ideas in this paper on the design of such a system are to construct an effective dissimilarity measure by including the direction of pen movement, to select a few representative signatures from the reference set by using a clustering procedure, and to decide the threshold as the maximum of dissimilarity measures among reference signatures multiplied by a constant. The effectiveness of the designed system is examined experimentally using a database provided by CADIX Co., Ltd. It consists of 2203 writings for 28 signatures which contain 10 in Latin letters (English alphabet) by foreigners, 14 in Japanese letters by Japanese, and 4 in Latin letters by Japanese. It comprises genuine signatures and forgeries. As a result of the experiment, the system turned out to work highly satisfactory: it achieved about 99% of an average correct verification rate. The inclusion of the direction of pen movement into the dessimilarity measure leads to about 3% of an average increase in the correct verification rate.

An automatic zip code recognition system for Japanese mail is proposed in this paper. It is assumed that a zip code is composed of three numerals and requited to be written in a specified frame. In actual images, however, the three numerals sometimes extend outside the specified frame and are not clearly separated. Considering this situation, the authors devised a system with two stages, the segmentation stage and the recognition stage. The segmentation stage consists of five steps: setting and adjusting of initial areas for numeral images (figures), calculation of the center of gravity of each figure, search for the horizontal and vertical boundaries of each figure, determination of the final area for each figure, and normalization of the figure in each final area. In the recognition stage, the Localized Arc Pattern Method (Arc method) proposed by Yoshimura et al. (1991) is implemented hierarchically; that is, a simple Arc method is applied first to each figure and a more complex one is applied subsequently unless the figure is identified in the first step. In the recognition process, every figure is judged as a numeral or otherwise rejected. The proposed system was applied to a database provided by the Institute for Post and Telecommunications Policy (IPTP). The segmentation algorithm yielded an adequate result. The recognition algorithm yielded scores as high as 90.6% in correct recognition rate and 0.7% in error rate. The best score of the precision index (P-index) specified by the IPTP was as low as 15.7 for the above mentioned IPTP database, while the score for another IPTP database was 16.9.