Tadahiro Azetsu

A self-organizing map with twin units capable of describing a nonlinear input-output relation applied to speech code vector mapping

This paper describes a new type of self-organizing map (SOM) with twin units as opposed to the single unit type conventional SOM proposed by Kohonen. The present self-organizing map with twin units (TW-SOM) can describe a nonlinear input-output relation with high accuracy. It is applied to voice conversion problem from bone conduction voice to air conduction voice (nonlinear code vector mapping), and its superiority over the conventional method using Linde-Buzo-Gray (LBG) algorithm is discussed. The tone quality of the converted voice is examined not only from the quantization distortion viewpoint, but also from the auditory sensation viewpoint through actual listening tests. The enhancement of the tone quality was experimentally confirmed.

Trilateral filter using rank order information of pixel value for mixed Gaussian and impulsive noise removal

The bilateral filter can remove Gaussian noise while preserving edges of the objects in an image. However the bilateral filter can not remove impulsive noise due to its edge preservation property. Therefore some robust bilateral filters have been proposed in order to also deal with impulsive noise. To make the bilateral filter more robust, this paper proposes a trilateral filter which incorporates a third weighting function based on the rank order information of pixel values into the bilateral filter. The effectiveness of the proposed method is verified in comparison with other conventional methods in experiments using the natural digital images corrupted by mixed Gaussian and impulsive noise.

Blind Separation and Sound Localization by Using Frequency-domain ICA

The independent component analysis (ICA) in the frequency domain is a method to deal with a blind signal separation problem in which propagation time delays are included in the mixing process of signals. We propose an extended method of the frequency-domain ICA accompanying the estimation of the relative propagation time delays and the propagation coefficient ratios. The effectiveness of the proposed method has been confirmed by simulation experiments. In addition, the sound localization by the proposed method is further discussed.

Gamma correction-based image enhancement for elderly vision

It is known that the lenticular transmittance of elderly people is lower than that of young people. Especially, it is remarkable for short wavelength lights. Therefore blue-tinged colors become darker comparing with the other colors in the field of vision of elderly people. In this paper, in order to improve the visibilities of blue-tinged colors in the elderly vision, the data-dependent gamma correction is proposed. Concretely, RGB components are converted by the gamma correction with a gamma which is set according to the hue, saturation and lightness contrast in each pixel. Through the experiments using some digital color images, the effectiveness of the proposed method is verified.

Application of peripheral auditory model to speaker identification

This paper discusses an approach for speaker identification using the multi-dimensional pulse signals generated from a model of a peripheral auditory system. The model of the peripheral auditory system employed here consists of a basilar membrane, hair cells, and auditory nerves. The input to this model is a speech signal divided into frames, and the outputs from which are the multi-dimensional pulse signals for each framed signal. The feature vectors based on the post-stimulus time histogram (PSTH) of the pulse signals are used for the speaker identification. Also, in order to improve the accuracy of the speaker identification, the feature vector conversion, using the mean and the diagonal matrix of standard deviations, is performed. The experiments were conducted for each Japanese vowel spoken by 12 speakers (9 males and 3 females), and the speaker identification accuracy is evaluated by 5 hold leave 2 out cross-validation for each vowel. The effectiveness of the proposed method has been verified by comparing with the conventional LPC analysis.

Lightness transform method considering visual feature of elderly person

It is known that the lenticular transmittance of elderly person is lower than that of young person. This low lenticular transmittance property causes the brightness decrease of the visual field of the elderly person. In this report, the lightness transform method of the image for the elderly person is proposed. In the proposed method, the colors, which are hard to see for the elderly person, are transformed into the colors, which are easy to see, by adding them the lightness contrast. The effectiveness of the proposed method is verified by some experiments.

Application of Subspace Method and Sparse Coding to Tissue Characterization of Coronary Plaque for High-speed Classification

Abstract The major cause of Acute Coronary Syndrome (ACS) is a rupture of coronary plaque. Therefore, the tissue characterization of coronary plaque is important for a diagnosis of ACS. In this study, we propose a method to use sparse features and its neighboring information obtained by a sparse coding. In the proposed method, the Radio Frequency (RF) signal obtained by the IntraVascular UltraSound (IVUS) method is expressed by a linear combination of the basis functions extracted from the learning signals by the sparse coding, and the code patterns of the expansion coeffcients of the basis functions are used for the tissue characterization. In addition, in order to perform a high-speed tissue characterization, the subspace method is employed as the classifier. The effectiveness of the proposed method has been verified by comparing the classification results of the proposed method with those of the frequency analysis-based conventional method applying to the data obtained from the human coronary arteries.

High performance hybrid-ICA to increase convergence speed and accuracy with use of RBF network

In this paper we first propose to use a radial basis function (RBF) network to increase the separation performance of blind signal separation (BSS). The independent component analysis (ICA) is often used for the BSS problem, but in general, the ICA employs the sigmoid function to describe the probability distribution of signal, more precisely the derivative of the logarithmic probability density function (PDF) of signal. In order to enhance the signal separation performance of BSS, we try to describe this nonlinear derivative function as accurately as possible by using RBF network. We further propose a hybrid ICA to make the most of the both advantages of the conventional ICA and the RBF based ICA. The proposed method is applied to several signal separation problems. The effectiveness of the proposed method has been confirmed by simulation experiments.