Norio Akamatsu’s research while affiliated with Tokushima University and other places

Conventionally, manual operations that specify positions of feature points such as eyes and nose are needed when morphing is carried out for a face image. In this work, the feature points are therefore extracted by using face area detection and a feature points decision methods to automate positional specification of feature points. As a result, the morphing of a face image can be carried out without manually specifying feature points. Face area detection is achieved by a threshold method using the YIQ color system. Feature points decision method extracts feature points by using a 3 layer perceptron type neural network (back-propagation). The attribute of the feature of eyes is defined to be a value of A in the color system LAB. In the same way, the attribute of feature points of the lip is defined as a value of B in the color system LAB. The extraction experiment of feature points was conducted from 120 face images by using the neural network, and the effectiveness of the present method was verified

Recently in medical fields, various imaging diagnostic technologies have been studied and used in practical. It is necessary to develop an automatic diagnosing processing system for detecting and diagnosing the internal organs. By the way, cardiac disease is one of the most common cause of death. Therefore, it is necessary to measure cardiac function quantitatively. The processing images are X-ray photograms of the left ventricle by cardiac catheterization. In this paper, we propose the detection system of asynergy in the left ventricle by using a neural network. Furtheremore, in order to demonstrate the effectiveness of the proposed method, we show the simulation example by using the real data.

In order to solve a stress problem, researchers have studied music therapy. It takes the therapist and patient a long time to select the music. Because the music used in music therapy is of various type. If the music for it is easily selectable, the music therapy can be carried out more effectively. In this paper, the purpose is extraction of features that may be influenced by the music. We pay attention to EEG (electroencephalogram) as an objective and absolute scale. In this paper, we propose a method that extracts features of the EEG by PCA (principal component analysis) and CDA (canonical discriminant analysis). Then we analyze each feature data by NN (neural network). In order to examine whether the proposal system is effective, we try computer simulations for the EEG classification. According to recognition rate by the NN, it was considered that the CDA extracted and classified the features of the EEG better than the PCA.

In this paper, we aim for construction of high-speed and high-accurate system using Fast Fourier Transform (FFT) for feature extraction, Simple-PCA (SPCA) for feature compression, and a neural network (NN) for recognition. In particular, we present a novel method based on Canonical Discriminant Analysis (CDA) to improve recognition accuracy for EMG. From results of computer simulation, it is shown that our approach is effective for improvement in recognition accuracy.

In this paper, a kidney region is extracted as a preprocessing of kidney disease detection. The kidney region is detected based on its contour information that is extracted from a CT image using a dynamic gray scale value refinement method based on the Q-learning. An initial point to extract the kidney contour is decided by training gray scale values along horizontal direction with Neural Network (NN). Furthermore the kidney contour is corrected by using the snakes more accurately. It is demonstrated that the proposed method can detect stably the kidney contour from CT images of any patients.

In baseball games, foul balls sometimes fly into the spectators seats. Some person can be seriously wounded for those balls. This research aims at developing a system which computes the orbit of a foul ball on real time using camera, predicts of the spot where foul balls fall and informs spectators about its danger. Detection of balls consists of the following three steps: At first each frame is processed using a smoothing filter for noise reduction. Next, temporal Laplacian is carried out for a series of those filtered images. Finally, whether a part of a ball or not is judged for each pixel by translating the resultant images into polar coordinates. The pitching motion is also considered. We report a method for prediction of foul ball falling spot with fuzzy inference.

This paper proposes a segmentation method of SAR (Synthetic Aperture Radar) images which uses a SOM(Self-Organizing Map). SAR images are obtained by observation using microwave sensor. They are segmented into the drift ice (thick, thin), and sea regions manually, and then features are extracted from partitioned data. However they are not necessarily effective for neural network learning because they can include incorrectly segmented data. Therefore, in particular, a multi-step SOM is used as a learning method to improve reliability of teacher data, and carries out classification. This process enable us to fix all mistook data and segment the SAR data using just data. The validity of this method was demonstrated by computer simulations using the actual SAR images.

There are some fundamental frequency detection methods for speech processing such as cepstrum analysis. However, the traditional methods need a lot of computing time and arithmetic processing. Moreover, a speech signal is converted into frequency domain by using an analysis section. In this paper, we propose a fast direct transformation (FDT). FDT extracts an amplitude feature of a signal by a simple computation. We perform the fundamental frequency detection of speech signal by using FDT. We compare the FDT algorithm with the conventional fundamental frequency detection methods by using teacher data (fundamental frequency) detected by the inspection. We perform an improvement as compared with an autocorrelation method by using FDT algorithm.

License plate recognition is very important technology for regulating an automobile society, However, doing so it is very difficult, because a background and a car surface color can be similar to that of the license plate. Furthermore, detecting license plates when cars are moving at a very high speed is difficult. In this paper, We propose a new method of extracting car license plates automatically by using a real-coded genetic algorithm(RGA). Using the proposed RGA, the most likely plate colors can be detected under various light conditions. First, the average brightness Y value of images are calculated. Next, the retation between the Y value and the most likely plate color thresholds upper and lower bounds are obtained with the RGA to estimate threshold equations. To prove the effectiveness of the proposed method, we describe simulations performed using real images.

We extract a kidney region as a preprocessing of kidney disease detection. The kidney region is detected based on its contour information that is extracted from a CT image using a dynamic gray scale value refinement method based on the Q-learning. An initial point to extract the kidney contour is decided by training gray scale values along horizontal direction with Neural Network (NN). Furthermore the kidney contour is corrected by using the snakes more accurately. It is demonstrated that the proposed method can detect stably the kidney contour from CT images of any patients.