Jun Tsukumo

Classification of handprinted Chinese characters using nonlinear normalization and correlation methods

A description is given of the classification of handprinted Chinese characters, using correlation methods for fast classification and a nonlinear normalization based on uniform relocation of the strokes used to form the character. Experimental results for handprinted Chinese character classification are presented. High classification capability, at 97.36% for the recognition rate and 99.44% for the rough classification was achieved for a large data set (ETL8), which includes 881 handprinted Chinese characters and 160 character patterns per character. A 94.42% recognition rate was achieved for a larger data set (ETL9B), which includes 2965 handprinted Chinese characters.<<ETX>>

On-line Japanese character recognition experiments by an off-line method based on normalization-cooperated feature extraction

It is shown that an offline character recognition method is effective for use in an online Japanese character recognition. Major conventional online recognition methods have restricted the number and the order of strokes. The offline method removes these restrictions, based on pattern matching of orientation feature patterns. It has been improved with developments in nonlinear shape normalization, nonlinear pattern matching, and the normalization-cooperated feature extraction method. It was used to examine 52,944 online Kanji characters in 1,064 categories. The recognition rate achieved 95.1%, and the cumulation recognition rate within the best five candidates was 99.3%.<<ETX>>

Handprinted Kanji character recognition based on flexible template matching

Describes a handprinted Kanji character recognition by hierarchical classification based on flexible template matching. In the proposed method, two kinds of flexible template matching are realized by nonlinear shape normalization and by nonlinear pattern matching, based on dynamic programming. The former has the efficiency for local shape restoration and the latter has the efficiency for local shape restoration in handprinted Kanji character variations.<<ETX>>

Development of a map vectorization method involving a shape reforming process

Described is an automatic map vectorization method to obtain geographical data from paper maps, for DB-systems using map information, such as GIS. The method consists of component separation, and skeleton vectorization. To obtain accurate vector data, several special segmentation techniques and a new skeleton vectorization technique is introduced in the method. The segmentation techniques are used to extract line drawings, character symbols, and other components, such as painted objects, from the original map image. The new skeleton vectorization technique involves a shape reforming process to reform vector shape distortion caused by raster-to-vector conversion. The process is based on the energy minimization principle. The process has an advantage in regard to reducing the various kinds of distortions, compared with using conventional processes. Through map input experiments with the method, it was proved that the method an obtain accurate vector data, compared with conventional techniques.<<ETX>>

A segmentation method for handwritten Japanese character lines based on transitional information

The authors propose a method to segment a character line image into individual character images and recognize them. To segment Japanese handwriting accurately, it is necessary to use character recognition results and contexts. However, recognition results might be wrong, or recognition confidence scores might be inaccurate. Dictionary consulting is not sufficient to deal with such ambiguous character recognition results. The paper reports on a method which hierarchically uses transitional information and a word dictionary for recognition results for all possible characters. Experimental results show that, for character line samples written roughly, 91.7% recognition rate is achieved while the recognition rate for a method without transitional information is 78.3%.<<ETX>>

Document Image Analysis For Reading Books

A fundamental problem in machine vision is to detect and identify special objects in an image. In the field of machine-reading for existing printed matter and books, a very important technique allows extracting and recognizing characters in desired text lines from a document image. This paper describes a hierarchical image segmentation, which separates a document image into its entities. Furthermore, a character segmentation, with minimum variance criterion, and a character recognition, based on three improved loci feature, have been developed as two elemental methods for reading books. In these experimental results using different commercial Japanese pocket books, 99% of text lines were correctly extracted. Also, it was successful in reading 99.30% of the Japanese characters and Chinese ideographs, as used in printed text.

A criterion for training reference vectors and improved vector quantization

In this paper, the criterion for training reference vectors is formulated in which the reference vectors are modified by the input vectors closer to decision boundaries. The authors present an improved vector quantization method, based on the above idea. Decision boundaries determined by this method are discussed and it is shown that the proposed method has several advantages as compared with conventional LVQ2. Experimental results for printed Japanese Hiragana characters recognition reveal that the proposed method is superior to LVQ2 and MLP in recognition ability.<<ETX>>

A multi-template learning method based on LVQ

A multitemplate learning method based on learning vector quantization (LVQ) is described. In this method, the learning process and the recognition process are carried out alternatively until all of the given data are recognized correctly with an increase in the number of reference vectors. The usefulness of the proposed method is demonstrated through preliminary simulations for artificial data and through recognition experiments for Japanese Hiragana characters compared with the k-means method and conventional LVQ. It is shown that better recognition results are obtained by the proposed method with fewer reference vectors than LVQ.<<ETX>>

Document Layout and Reading Sequence Analysis by Extended Split Detection Method

This paper describes an Extended Split Detection Method that can hierarchically segment a machine-printed page image with a complex layout into smaller layout elements. The method performs piecewise-linear segmentation using many kinds of separator elements such as field separators, lines, edges of figures, and edges of white background areas. Furthermore, this method represents an analyzed layout of a hierarchical structure in a tree data structure, in which all nodes are traversed according to the simple rules for generating the reading sequence. We demonstrated that the new method increases the correct character line segmentation rate by 15.5%, to 95.5%, and we achieved a correct reading sequence generation of 88.1%.

Shape based learning for a multi-template method, and its application to handprinted numeral recognition

Character recognition using multi-template methods is promising. Higher classification performance can be achieved according to an increase in the number of templates. However, classification performance is saturated because there is classifiability loss in feature extraction. The paper proposes a new multi-template method which learns training patterns with character shape information assigned by the authors. This method uses contour feature and direction feature, and includes a character shape consistency test applied to the conventional multi-template methods. The paper presents experimental results obtained from handprinted numerals. On the ETL-6 database classification experiment the classification rate was 99.19% and the substitution rate was 0.03%. A higher classification rate could be achieved.