M. Fukumi’s research while affiliated with Tokushima University and other places

In this paper, we present a character recognition method using features extracted from a star layered histogram and trained using neural networks. After several image preprocessing steps, the character region is extracted. Its contour is then used to determine the center of gravity (COG). This CoG point is used as the origin to create a histogram using equally spaced lines extending from the CoG to the contour. The first point the line touches the character represents the first layer of the histogram. If the line extension has not reached the region boundary, the next hit represents the second layer of the histogram. This process is repeated until the line touches the boundary of the character's region. After normalization, these features are used to train a neural network. This method achieves an accuracy of about 93% using the MNIST database of handwritten digits.

Many car accidents are caused by driver's deviation from normal condition like carelessness. We aim to construct a driving assist system that is able to detect driver's deviation signal from normal condition. The system detects the deviation signal using driver's time-series head motion information. In this paper, we optimize categorization of driver's head motion using two kinds of unsupervised neural networks: Self-Organizing Maps and Fuzzy Adaptive Resonance Theory, and discuss the relation between vigilance parameter and integrated categories.

Many car accidents are caused by driver's deviation from normal condition like carelessness. We aim to construct a driving assist system that is able to detect driver's deviation signal from normal condition. The system detects the deviation signal using driver's time-series head motion information. In this paper, we analyze driving movies taken by monocular in-vehicle camera, and examine driver's head position category in safety verification at intersections for quantification of head motion information. Moreover, we propose a quantifiable categorizing algorithm of head motion using two kinds of unsupervised neural networks, and it provides a possibility of quantification of the head position.

In pattern recognition, the principal component analysis (PCA) is one of the most famous feature extraction methods for dimensionality reduction of high-dimensional datasets. Furthermore, Simple-PCA (SPCA) which is a faster version of the PCA, has been carried out effectively by iterative operated learning. However, in SPCA, when input data are distributed in a complex way, SPCA might not be efficient because it is learned without class information of the dataset. Thus, SPCA cannot be said that it is optimal for classification. In this paper, we propose a new learning algorithm, which is learned with the class information of the dataset. Eigenvectors spanning eigenspace of the dataset are obtained by calculation of data variations belonging to each class. We will show the derivation of the proposed algorithm and demonstrate some experiments to compare the SPCA with the proposed algorithm by using UCI datasets.

In recent years, personal authentications using biological information are used for protection of personal data and confidential information in local governments and companies. In this paper, we propose a novel personal identification method using footsteps. The users' mental burdens of the proposal technique is a little because the footsteps can be easily recorded without special equipments. First, the proposed method detects footstep sections from the recorded signals. Then, the acoustic feature parameters, which are Mel-Frequency Cepstral Coefficients (MFCCs), ΔMFCCs, and ΔLogarithm Powers (ΔLPs), are extracted as footstep features from the footstep section. Finally, persons are identified by k-Nearest Neighbor (k-NN) in which Dynamic Programming matching algorithm (DP) is used as a distance measure and/or Gaussian Mixture Models (GMMs). We conduct personal identification experiments using 720 footstep data which are recorded from 12 test subjects for evaluating the proposed method. From the experimental results, average accuracies of overall footwear are 79.9% and 92.8% in k-NN and GMMs.

This paper presents a novel method for optimal region-of-interest (ROI) segmentation for palmprint feature extraction based on the largest inscribed circle. This ensures that the optimal amount of features can be extracted, unlike square-based methods which exclude a substantial area on the outside region of the palmprint image. After position normalization, the middle portion of the palmprint is searched to determine the center from which the largest inscribed circle can be extracted. The circular area is then unwrapped into a fixed-size rectangular strip which is further preprocessed to remove redundancies and then split into seven equal square sub-images. A layered approach is then adopted during the matching stage where each square is successively matched and polled to produce a matching score. Experiments are performed using the `PolyU Palmprint Database' and results show that this is a viable method for palmprint feature extraction, with a recognition rate of above 90% obtained.

In this paper, we propose a classification system for fingerprint images that is based on the number of registered fingerprint persons. Most automated fingerprint identification systems use prior classification of fingerprint for improvement of efficiency verification using minutiae as features. However, methods that use fingerprint minutiae needs improvement because they are limited to the number of classable data. Therefore, many fingerprints are classified together, consequently taking a long time to match and verify a given fingerprint. In this work, we propose a system that classifies fingerprint patterns into individual classes. Instead of the classification using minutiae, we propose a classification system that is based on individual features and the number of registered persons. Efficiency verification improves because we donpsilat need to compare an input fingerprint image to all registered fingerprint images using this system. The proposed system carries out classification using neural network.

In this paper, new statistical learning algorithms with kernel function are presented. Recently, iterative learning algorithms for obtaining eigenvectors in the principal component analysis (PCA) have been presented in the field of pattern recognition and neural network. However, the Fisher linear discriminant analysis (FLDA) has been used in many fields, especially face image analysis. The drawback of FLDA is a long computational time based on a large-sized covariance matrix and the issue that the within-class covariance matrix is usually singular. In order to overcome this difficulty, we proposed the feature generation method Simple-FLDA which is approximately derived from geometrical interpretation of FLDA. This algorithm is similar to Simple-PCA and does not need matrix operation. In this paper, new statistical kernel based learning algorithms are presented. They are extended versions of Simple-PCA and Simple-FLDA to nonlinear space using the kernel function. Their preliminary simulation results are given for a simple face recognition problem.

The purpose of this study is to develop an useful interactive interface to operate a welfare apparatus.Moreover, an interactive operation screen is proposed, which is based on the presumptions of some physiological knowledge and Evolutionary Eye Sensing. The proposed system uses a noncontact type interface with the Evolutionary Eye Sensing. A face is tracked and eye region image is captured automatically by a controlled pan-tilt-zoom camera. Then, an iris is tracked and eye movement is measured. An operation screen is divided into 9 areas. A user can select these 8 areas by the eye movement and decide by the eye fixation. This proposed system is not necessary a special calibration. The effectiveness of the proposed system is evaluated by experiments with 10 subjects. The results indicate that the proposed system is easy to use for the beginner, and the user can be proficient in operation by exercises.

In this paper, we propose a method for detecting the mood much music for prefrontal cortex electroencephalogram (EEG) activity. The analyzed EEG frequencies contain significant and immaterial information components. We focused on the combinations of the significant frequency. These frequency combinations are thought to express personal features of EEG activity. In the proposed method, we calculate the spectrum of these frequency combinations rates that does not include the noise frequency components and evaluates whether the music matches the user's mood through a simple threshold processing. Then, a genetic algorithm (GA) is used to specify the frequency of personal features on the EEG. The threshold vale used the threshold processing is the average value of the spectrum rates specified EEG frequency combinations. Finally, the performance of the proposed method is evaluated using real EEG data.