Kazumi Odaka

On-line handwriting recognition

For large-alphabet languages, like Japanese, handwriting input using an online recognition technique is essential for input accuracy and speed. However, there are serious problems that prevent high recognition accuracy of unconstrained handwriting. First, the thousands of ideographic Japanese characters of Chinese origin (called Kanji) can be written with wide variations in the number and order of strokes and significant shape distortions. Also, writing box-free recognition of characters is required to create a better man-machine interface. Intense research performed over the past 15 years to answer the most pressing recognition problems is described. Prototype systems are also described. The man-machine interfaces made possible by online handwriting recognition and anticipated advances in both hardware and software are discussed. >

A prototype system for interpreting hand-sketched floor plans

A prototype system for drawing interpretation that automatically converts hand-sketched floor-plans into the CAD format is presented. The main problems in interpreting hand-sketched drawings are positional ambiguity and shape distortion. In the proposed method, positional ambiguity is able to be corrected in a nearest grid-location manner because floor-plan elements are normally drawn on/along grid lines that correspond to the modular proportions. Shape distortion is able to be corrected by template-matching of closed regions because architectural elements mostly consist of one or more closed regions of basic figures: triangles, squares, circles, and so on. The system was tested using 150 floor-plan drawings hand-sketched by three subjects without the use of rulers or templates. The results showed that sufficient recognition performance for practical use was able to be obtained.

On-line cursive Kanji character recognition as stroke correspondence problem

This paper describes a stroke-number and stroke-order free on-line Kanji character recognition method by a joint use of two complementary algorithms of optimal stroke correspondence determination: one dissolves excessive mapping and the other dissolves deficient mapping. Also, three kinds of inter-stroke distances are devised to deal with stroke concatenation or splitting and heavy shape distortion. Only a single reference pattern for each of 2,980 Kanji character categories is generated by using training data composed of 120 patterns written with the correct stroke-number and stroke-order. Recognition tests are made using the training data and two kinds of resting 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.

Form processing based on background region analysis

We present a novel approach for processing form documents based on background region analysis. Our goal is to achieve line-property-free form processing. Background regions can be extracted independently of line width or length, and multi-layer analysis employing a series of coarse-to-fine background images makes it possible to extract background regions regardless of small line-breaks. We propose two multi-layer analysis algorithms for different situations. One is applied in a registration process of a form model. It reliably extracts box regions from un-filled forms without using any model. The other is applied in a character extraction process. By using a spatial model of a form, it reliably extracts background regions, and re-integrates these regions if they are divided by characters written in the boxes. From these re-integrated regions, the exact locations of the character boxes are determined on the input image. Besides these algorithms, we present a form identification method that uses coarse background images. We implemented the algorithms into a prototype system that processes pre-printed forms. 50 types of existing forms were tested without any customization. Model registration, character extraction, and form identification were reliably carried out.

On-line cursive Kanji character recognition using stroke-based affine transformation

This paper describes a distortion-tolerant online Kanji character recognition method using stroke-based affine transformation (SAT). The first part of the method determines one-to-one stroke correspondence between an input pattern and each reference pattern. The second part applies optimal SAT to each stroke of the input pattern to absorb handwriting distortion. The last part calculates the inter-pattern distance between the reference pattern and the SAT-superimposed input pattern. Only a single reference pattern for each of 2,980 Kanji character categories is generated by using training data written carefully with the correct stroke-number and stroke-order. Recognition tests are made using two kinds of test data in the square style and in the cursive style written by 36 different people; recognition rates of 98.4% and 96.0% are obtained.

Adaptive normalization of handwritten characters using global/local affine transformation

Conventional normalization methods for handwritten characters have limitations, such as preprocessing operations because they are category-independent. The paper introduces an adaptive or category-dependent normalization method that normalizes an input pattern against each reference pattern using global/local affine transformation (GAT/LAT) in a hierarchical manner as a general deformation model. Experiments using input patterns of 3171 character categories, including Kanji, Kana, and alphanumerics, written by 36 people in the cursive style against square style reference patterns show not only that the proposed method can absorb a fair large amount of handwriting fluctuation within the same category, but also that discrimination ability is greatly improved by the suppression of excessive normalization against similarly shaped but different categories.

Discrepancy between brain magnetic fields elicited by pattern and luminance stimulations in the fovea: Adequate stimulus positions and a measure of discrepancy

SummaryA conventional equivalent current dipole estimation provides one of the quantitative measures to evaluate the discrepancy between two single-dipole-like magnetic field patterns, though there is one problem; all stimulus positions in the visual field do not necessarily contribute to the generation of a single-dipole-like magnetic field. Another important problem occurs when the field pattern is complex and cannot be approximated by a dipole. This makes it difficult to evaluate the discrepancy between two magnetic field patterns by the dipole parameters. In this paper, we determined the stimulus positions adequate for generating single-dipole-like magnetic field patterns by evaluating the magnetic fields goodness-of-fit to the field generated by a single dipole. We propose to use a similarity (SIM) as a quantitative measure of the discrepancy between two complex magnetic field patterns. The SIM is defined as an angle between two magnetic field vectors. We evaluated the discrepancy between the 100 ms post-stimulus responses to pattern-reversal (Rv) stimulus, pattern-onset (Pat) stimulus, and luminance-onset (Lumi) stimulus. The following results were obtained: (1) Stimulation of some of the octants in the fovea, far from the vertical meridian, elicited a single-dipole-like magnetic field pattern at a latency of 100 ms, though stimulation of the central part of the fovea, and stimulation of the octants along the vertical meridian, did not elicit a single-dipole-like magnetic field pattern; (2) The discrepancy between responses was quantitatively evaluated by the SIM even if the field patterns were complex; (3) The SIM analysis showed that the discrepancy between the responses to the Rv and the Lumi stimuli, as well as that between the responses to the Pat and the Lumi stimuli, were greater than that between the responses to the Rv and the Pat stimuli.

Analysis and evaluation of dictionary learning on a handy type pen-input interface for personal use

We present the learning characteristics of personal dictionaries used for handy type pen-input interfaces. The personal dictionary grows by acquiring misrecognized input patterns whenever they are generated. Personal character pattern data is collected using a handy type tablet. The correct recognition rate of the data, which consists of 756 characters with 391 different characters, is improved from 79.9% to 96.2% for data used in learning, and from 78.0% to 95.4% for unlearnt data. On the contrary, the decrease of recognition rate by this learning is 0.1% –0.2%, which is negligible. The results show that the learning yields a handy type pen-input interface that is sufficient for practical use, moreover, this paper clarifies the learning characteristics against changesin writing condition, the number of objective categories, the number of registered patterns in the personal dictionary, and the formation process of the personal dictionary. Analysis and evaluation of the results gives useful instruction on the design of recognition systems.

Combining statistical pattern recognition approach with neural networks for recognition of large-set categories

We present a two-stage hierarchical system consisting of a statistical pattern recognition (SPR) module and artificial neural network (ANN) to recognize a large number of categories including similar category sets. In the first stage, the SPR module performs classification. If the first candidate does not belong to a pre-determined similar category set, the first candidate is accepted as the final result; otherwise, the first candidate is sent to the ANN module. In the second stage, ANN performs classification for similar categories to select a correct candidate from the predetermined candidate set designated by the first candidate. The new scheme offers improved system performance by sharing tasks between SPR and ANN according to the degree of classification difficulty and forming specialized ANNs for each similar category. The system achieves higher performance for the recognition of 3,201 handprinted characters than a traditional system constructed with just the SPR module.

Development of handwritten character input interface for multimedia terminal and its applications

In a practical online character-recognition technique for the character input of a multimedia terminal, it is important to obtain a highly accurate real-time interface for handwritten characters. This paper describes a system that includes: (1) a method of character recognition tolerant to variations of handwriting; (2) fine discrimination logic for characters; (3) character recognition using stroke-end matching; (4) dictionary learning and algorithms for particular users; and (5) a front-end processor requiring fewer manual operations and having independence and adaptability for applications. The proposed method has been applied to a real medical reservation service, and its usefulness has been confirmed.