This paper proposes a method to improve the off-line character classifiers which are learned from examples by using the virtual examples synthesized from on-line character database. To obtain the good classifiers, usually a large database which contains enough number of variations of handwritten characters is required. However, in practice collecting enough number of data is time-consuming and costly. In this paper, we propose a method to train SVM for off-line character recognition based on the artificially augmented examples using on-line characters. In our method, a virtual example is synthesized from an on-line character by applying affine transformation to each stroke. SVM classifiers are trained by using the artificially generated patterns. To reject inappropriate artificial examples, a preliminary SVM is learned from the original set of samples, and then used for data selection. Using the augmented training samples, the final SVM is obtained. We examine the effectiveness of the proposed method by experiments of handwritten Japanese Hiragana character classification.

This paper presents a digit classification system to recognize telephone numbers written on signboards. Candidate regions of digits are extracted from an image through edge extraction, enhancement and labeling. Since the digits in the images often have skew and slant, the digits are recognized after the skew and slant correction. To correct the skew, Hough transform is used, and the slant is corrected using the method of circumscribing digits with tilted rectangles. In experiments, we tested a total of 1332 images of signboards with 11939 digits. We obtained a digit extraction rate of 99.2% and a correct digit recognition rate of 98.8%.

The purpose of our research is to improve the recognition rate of an off-line handwritten character recognition system using HMM (hidden Markov model), so that we can use the system for practical application. Due to the insufficient recognition rate of ID HMM character recognition systems and the requirement for a huge number of learning samples to construct 2D HMM character recognition systems, HMM-based character recognition systems have not yet achieved sufficient recognition performance for practical use. In this research, we propose the character recognition method that integrates 4 simply structured 1D HMMs all of which are based on feature extraction using linear filters. The results of our evaluation experiment using the Hand-Printed Character Database (ETL6) showed that the first rank recognition rate of the test samples was 98.5% and that the cumulative recognition rate of top 3 candidates was 99.3%. Although our method is relatively easy to implement, it can work even better than 2D HMM method. These results show the proposed method is very effective.

Discusses the relation between technical seeds in character recognition and market needs. The seeds can meet the market needs providing that the seeds are at a high enough level. The character recognition techniques used so far however, have not matured owing to many constraints. In other words, human beings have made many artificial limitations so as to put premature OCR techniques into practical use. The technical progress and competition are making these constraints unnecessary. In this context, unconstrained OCR techniques are explained.

The purpose of our research is to improve the recognition rate of offline character recognition systems using the HMM (Hidden Markov Model) without increasing the number of HMM parameters too much. Some 2-dimensional HMM character recognition systems have been proposed to increase representational power. However since 2D HMM has a much more complex structure and thus requires much more parameters than 1-dimensional HMM, it becomes very hard to gather sufficient samples in order to guarantee successful generalization. To overcome the problem, we propose a method for character recognition using 1D HMMs in multiple directions with 2-dimensional feature extraction. To further improve the performance, some voting methods using a bagging algorithm are also exploited. In our experiment, the recognition rate is increased by about 1% with the multiple directional HMM character recognition system compared to the 1D HMM character recognition system. The recognition rate is further increased by about 1% with the HMM character recognition system using a bagging algorithm.

In this paper we propose a method for face recognition from an off-frontal input image. Our method is based on the bidirectional view synthesis. Firstly, from an off-frontal input image, a frontal "virtual" view is synthesized based on the linear combination model. Then, pattern matching is performed to select the possible matches. If several candidates for the matching remain, the backward view synthesis, from each candidate to the input view, is carried out. A reliable result is obtained by this bidirectional view synthesis method.

In an automatic music score recognition system, it is very important to extract heads and stems of notes, since these symbols are most ubiquitous in a score and musically important. The purpose of our system is to present an accurate and high-speed extraction of note heads (except the whole notes) and stems according to the following procedure. (1) We extract all regions which are considered as candidates of stems or heads. (2) To identify heads from the candidates, we use a three-layer neural network. (3) The weights for the network are learned by the back propagation method. In the learning, the network learns the spatial constraints between heads and surroundings rather than the shapes of heads. (4) After the learning process is completed we use this network to identify a number of test head candidates (5) The stem candidates touching the detected heads are extracted as true stems. As an experimental result, we obtained high recognition rates of 99.0% and 99.2% for stems and note heads, respectively. It took between 40 to 100 seconds to process a printed piano score on A4 sheet using a workstation. Therefore, our system can analyze it at least 10 times as fast as manual methods.

This paper proposed a new method for the face extraction from a complex background. Based on the space gray level dependence (SGLD) matrices, the facial texture model composed by a set of inequalities was derived. Using this textural model, the authors designed a kind of scanning scheme for face detection in the complex backgrounds. An experiment using 60 images containing 150 faces gave the error rate 0% with the false alarm rate 5.4%.

A new, high speed method filling in the contours of alpha-numeric characters which produces accurate and correct image patterns is proposed. The method is called the improved edge fill method because it improves on a previously developed edge fill method. It does so by selecting fill pixels from contour lines expressed by Freeman's chain code and using combinations for neighboring the codes and accordingly processing time is reduced by about 20% over conventional methods. For binary image contour filling, the method eliminates faults found in conventional edge fill methods, and in consequence the areas inside contour lines are filled in rapidly and correctly. The method's effectiveness is examined in experiments using both Arabic numerals and letters of the Roman alphabet. As a result it is found that this filling method is able to produce correct and accurate image patterns and that it can be applied to alpha-numeric character contour filling.>

A pattern-oriented segmentation method for optical character recognition that leads to document structure analysis is presented. As a first example, segmentation of handwritten numerals that touch are treated. Connected pattern components are extracted, and spatial interrelations between components are measured and grouped into meaningful character patterns. Stroke shapes are analyzed and a method of finding the touching positions that separates about 95% of connected numerals correctly is described. Ambiguities are handled by multiple hypotheses and verification by recognition. An extended form of pattern-oriented segmentation, tabular form recognition, is considered. Images of tabular forms are analyzed, and frames in the tabular structure are extracted. By identifying semantic relationships between label frames and data frames, information on the form can be properly recognized.>

An algorithm to normalize the skew of document images is proposed. The skew angle is detected in two stages. In the first stage, connected regions in an image are extracted and some feature parameters are extracted for each region. In the second stage, the Hough transform is calculated for the parameters, and the angle which gives the minimum of the transform is estimated as the skew angle. In experiments using CCITT standard documents, a detection accuracy of less than 0.1 degrees is obtained for printed documents. When graphical elements are included in the documents in addition to printed characters, the accuracy deteriorates to 0.2 degrees .>