Fumio Yoda

Field extraction method from existing forms transmitted by facsimile

In this paper, we present a field extraction method in form reading application, for existing forms transmitted by facsimile. In faxed form images, it is difficult to determine field locations exactly due to distortions on the form images. To overcome this problem, exclusive forms have been used. In order to make the field extraction easy, these exclusive forms have restrictions on the style. However, existing forms have various styles and no restrictions. Thus, it is hard to read faxed existing forms. The proposed method corrects the location of fields by estimating the geometric transformation of the form image based on the pairs of corresponding feature-points extracted on the form image and a form model. These feature-points can be extracted from table lines, underlines, character strings, and illustrations on existing forms. Moreover, the method estimates the geometric transformation with robustness for the illegal pairs of corresponding feature-points. These characteristics allow the method to read existing forms transmitted by fax. The experimental results with 50 different types of existing forms reveal the effectiveness of the method.

Structural character recognition using simulated annealing

The authors propose a structural character recognition method using simulated annealing, where the dissimilarity between an input and a reference pattern is evaluated by matching stroke and loop objects extracted from these patterns. This method can compute the matching at a high-speed. It can also avoid an incorrect matching problem when a object is divided into several objects in a broken character, by generating a hypothesis for an objects integration in the matching process. They have applied this method to handwritten digit recognition. The experimental results reveal that this method has the same degree of recognition rate and about 1/10 of the computing time compared with a conventional matching method using probabilistic relaxation. Moreover, generating a hypothesis for an objects integration is effective for dealing with broken characters.

Recognition of handprinted Kanji characters by the stroke matching method

This paper describes an improved stroke matching method for the recognition of handprinted Kanji characters. Using the stroke feature and two additional global features, a recognition rate of over 90% has been obtained for about 1000 Kanji characters.