When recognizing handwritten characters, normalization of characters has been usually done prior to feature extraction. Some authors point out that improvement of normalization increases the recognition rate. Thus, normalization techniques must be carefully examined. We propose a Gabor filter based handwritten character recognition system. Our main interest is to select the best normalization technique for Gabor filters. We conduct comparative experiments and show that position, slant, and thickness normalization is best, with regard to minimizing the error.

A bootstrap technique for nearest neighbor classifier design is proposed. Our primary interest in designing a classifier is in small training sample size situations. Conventional bootstrapping techniques sample the training samples with replacement. On the other hand, our technique generates bootstrap samples by locally combining original training samples. The nearest neighbor classifier is designed on the bootstrap samples and is tested on the test samples independent of training samples. The performance of the proposed classifier is demonstrated on three artificial data sets and one real data set. Experimental results show that the nearest neighbor classifier designed on the bootstrap samples outperforms the conventional k-NN classifiers as well as the edited 1-NN classifiers, particularly in high dimensions.

An anti-regularization technique which has been recently proposed by Raudys (1995) is studied in small training sample size situations. Experimental results show that as long as the weights of a network are initialized in a very narrow interval, the anti-regularization technique offers significant advantages in terms of both the generalization ability and learning time.

We study a Gabor filter-based feature extraction method for handwritten numeral character recognition. The performance of the Gabor filter-based method is demonstrated on the ETL-1 database. Experimental results suggest that the Gabor filter-based method should be considered in recognition of handwritten numeric characters.

It is widely believed in the pattern recognition field that when a fixed number of training samples is used to design a classifier, the generalization error of the classifier tends to increase as the number of features gets larger. In this paper, we discuss the generalization error of the artificial neural network (ANN) classifiers in high-dimensional spaces, under a practical condition that the ratio of the training sample size to the dimensionality is small. Experimental results show that the generalization error of ANN classifiers seems much less sensitive to the feature size than 1-NN, Parzen and quadratic classifiers.

A method for handprinted Chinese character recognition based on Gabor filters is proposed. The Gabor approach to character recognition is intuitively appealing because it is inspired by a multi-channel filtering theory for processing visual information in the early stages of the human visual system. The performance of a character recognition system using Gabor features is demonstrated on the ETL-8 character set. Mental results show that the Gabor features yielded an error rate of 2.4% versus the error rate of 4.4% obtained by using a popular feature extraction method.

Distortion invariant pattern recognition is an interesting problem from the biological and technological point of view. However, it has not yet been solved by neural networks in satisfactory way. This paper investigates an associative neural network system to improve the recalling accuracy for distortion patterns. On a perception type of neural network with feedback, error back-propagation algorithm and energy function are used for a learning process and a recalling process, respectively. By using gradated patterns as learning and unknown patterns, it is shown that the recalling accuracy becomes higher than using original pattern themselves.

The modified Fisher criterion takes into consideration the difference of covariance matrices. It is shown from experimental results that the proposed method is superior to the orthonormal discriminant vector and the Fehlauer-Eisenstein methods in terms of the error probability.>

In general, a two-dimensional display is defined by two orthogonal unit vectors. In developing the display, discriminant analysis has the shortcoming that, in general, the extracted axes are not orthogonal. In order to overcome this shortcoming, the authors propose a discriminant analysis which provides an orthonormal system in the transformed space. The transformation preserves the discriminatory ability in terms of the Fisher ratio. The authors present a necessary and sufficient condition under which discriminant analysis in the original space provides an orthonormal system. Relationships between orthogonal discriminant analysis and the Karhunen-Loeve expansion in the original space are investigated.>

The learning behavior is investigated. The game has three variable-structure stochastic automata and a random environment. In the game the players do not possess prior information concerning the payoff matrix, and at the end of every play all the players update their own strategies on the basis of the response from the random environment. Under such situations, if a payoff matrix satisfies some conditions, it can be shown that the learning behavior of the automata converges to the optimal strategies.

A new method is proposed for realizing a stochastic system from interconnected smaller stochastic systems. Each smaller stochastic system considered is composed of two states. The basic tool is the partition with a substitution property which exists on the transition matrix of the interconnected stochastic system. Using the authors' method, a stochastic system withnstates is decomposed intom(m≥n-1) component systems where each system has two states. The following results are derived: for a given outer input, each component system hasn-1 state transition matrices; a relationship exists among some entries of transition matrices; and the whole system is composed ofmcomponent stochastic systems, butm-n+1 component systems are selected freely.