Toyohisa Kaneko

Straight-Line Approximation for the Boundary of the Left Ventricular Chamber from a Cardiac Cineangiogram

A cardiac cineangiogram is an X-ray motion picture of the heart. Boundary detection is an essential task to compute valuable quantitative information such as the temporal volume change and the wall velocity. Previously, boundary detection methods treated frames independent of each other. Consecutive frames of a motion picture are, however, highly correlated and therefore this property can be exploited. The method described takes advantage of the known location of the boundary on previous frames. The previous boundary is divided into a set of segments along which local rectangular regions are set up on the present frame. This boundary is approximated by a set of straight lines which minimize the square error in each rectangular region with the spatial derivative as its weight. The method is a fast computing algorithm and relatively insensitive to white noise because the resultant positive and negative weights tend to cancel this effect. An experiment was carried out, and good agreement with boundaries detected by humans was obtained.

Effect of Coefficient Rounding in Floating-Point Digital Filters

The effect of coefficient rounding in digital filters using floating-point arithmetic is analyzed. An error is defined as the difference between the output of a digital filter realized with a finite number of bits and the output of the same filter using infinite number of bits. It is shown that this error consists of two uncorrelated components. One component is due to the rounding of the arithmetic operations alone, and the other component represents the additional error due to coefficient rounding. Expressions are derived for the mean-square value of the error for the three canonical programming forms. The stability problem is studied by examining the shift of the pole locations due to the coefficient rounding. The mean-square value of the shift is calculated. Good agreement between theoretical and experimental results is obtained.

Speaker-independent isolated word recognition using label histograms

In this paper, a simple and fast method for speaker-independent isolated word recognition is presented. This method is regarded as simplification of the approach based on the Hidden Markov Model (HMM). In the proposed method, all training and decoding data are transformed into label strings by vector quantization. By segmenting the label strings of utterances into N pieces with equal duration, label histograms are computed in the training mode. In recognition, the label string of an input word is also divided into equal N segments, and the likelihood is computed with the corresponding histogram. It will be shown that the computational cost of this method is relatively low. This method is applied to the recognition of 32-Japanese-word vocabulary, and achieved a recognition accuracy comparable to or better than that of the conventional approaches.

Optimal task switching policy for a multilevel storage system

Capacity demands for computer memory are increasing. A multilevel storage system provides an economically feasible solution without seriously affecting the total response time. An M-level storage system is considered in this paper. The capability of a digital computer with a multilexel storage system is best enhanced in a multiprogramming environment. In a high level storage system, determination of a best task switching policy becomes an important consideration. In this paper a queuing network is introduced to describe distribution and flow of tasks in the system. An optimal switching policy is determined in relation to the systems overhead time. It is shown that in heavily CPU-limited cases the determination becomes a very simple one; namely, the best policy is given as the threshold level at which the accumulation of the average access time exceeds the overhead time.

Speaker-independent isolated word recognition using n-segment label histogram method

A simplified version of a hidden Markov model (HMM), referred to as the N-segment label histogram (NLH) method, is proposed for speaker-independent isolated word recognition. The NLH method can be considered as an HMM with a uniform duration for each state. It can also be treated as a statistical pattern recognition approach using linear compander and probabilistic measures. During the training, the label histograms are computed for N equal segments of the input. The probability associated with each label is computed after normalization. During the recognition, the input is partitioned into N equal segments, the corresponding label that maximizes the label probability of the input word determines the recognition result. Since the NLH method requires only about one-tenth the computation of the HMM method, it is more suitable for implementation on small computers. Furthermore, it does not require alignment along the time axis as do the HMM and DP matching techniques. The utterance fluctuation along the time axis is handled statistically and is applicable to a large data set. Experimental results indicated that the NLH method rendered almost the same recognition rate as the other two methods while requiring much less computation. The linear approximation of the likelihood function enables the implementation of the proposed algorithm on the IBM PC/AT for real time speech recognition.