In computer graphics since objects atre constructed by lines and curves, the high-speed curve generator is indispensible for computer aided design and simulatation. Since the functions of graphic generation can be represented as a series of matrix operations, in this paper, two kind of the high-speed Bezier curve generator that uses matrix equation and a recursive relation for Bezier polynomials are designed. And B-spline curve generator is designed using interdependence of B-spline blending functions. As the result of the comparison of designed curve generator and reference [5], [6] in the operation time and number of operators, the curve generator with 1-dimensional systolic array processor for matrix vector operation that uses matrix equation for Bezier curve is more effective.

A two dimensional constant geometry FFT algorithms and architectures with shuffled inputs and normally ordered outputs are presented. It is suitable for VLSI implementation because all buterfly stages have identical, regular structure. Also a methodology using shuffled FFT inputs and outputs to halve the number of butterfly stages connected by a global interconnection which requires much area is presented. These algorithms can be obtained by shuffling the row and column of a decomposed FFT matrix which corresponds to one butterfly stage. Using non-recursive and recursive pipeline, the degree of serialism and parallelism in FFT computation can be adjusted. To implement high performance high radix FFT easily and reduce the amount of interconnections between stages, the method to build a high radix PE with lower radix PE 's is discussed. Finally the performances of the present architectures are evaluated and compared.

Phase retrieval is concerned with the reconstruction of a signal from its Fourier transform magnitude (or intensity), which arises in many areas such as X-ray crystallography, optics, astronomy, or digital signal processing. In such areas, the Fourier transform phase of the desired signal is lost while measuring Fourier transform magnitude (F.T.M.). However, if a reference 'signal is added to the desired signal, then, in the Fourier trans form magnitude of the added signal, the Fourier transform phase of the desired signal is encoded. This paper addresses uniqueness and retrieval of the encoded Fourier phase of a multidimensional signal from the Fourier transform magnitude of the added signal along with the Fourier transform magnitude of the desired signal and the information of the additive reference signal. In Part I, several conditions under which the desired signal can be uniquely specified from the two Fourier transform magnitudes and the additive reference signal are presented. In Part II, the development of non-iterative algorithms and an iterative algorithm that may be used to reconstruct the desired signal(s) is considered.

Phase retrieval is concerned with the reconstruction of a signal from its Fourier transform magnitude (or intensity), which arises in many areas such as X-ray crystallography, optics, astronomy, or digital signal processing In such areas, the Fourier transform phase of the desired signal is lost while measuring Fourier transform magnitude (F.T.M.). However, if a reference 'signal is added to the desired signal, then, in the Fourier trans form magnitude of the added signal, the Fourier transform phase of the desired signal is encoded This paper addresses uniqueness and retrieval of the encoded Fourier phase of a multidimensional signal from the Fourier transform magnitude of the added signal along with Fourier transform magnitude of the desired signal and the information of the additive reference signal In Part I, several conditions under which the desired signal can be uniquely specified from the two Fourier transform magnitudes and the additive reference signal are presented In Part II, the development of non-iterative algorithms and an iterative algorithm that may be used to reconstruct the desired signal (s) is considered

In this paper we consider efficient 1-D and 2-D linear-phased half-band filter designs. We first introduce a new derivation of the existing Vaidynathan-Nguyen 's half-band filter design method, which verifies that the design provides optimal half-band filters. We then propose an approximately-linear-phased IIR half-band filter design method, which is based on the all-pass equalizer design with the linear phase -$\omega$/2. Finally, we propose an efficient method to design optimal 2-D half-band filters, for which we utilize a 2-D all pass prototype filter of half the order of the desired 2-D half-band filters.

In the non-destructive evaluation techniques using ultrasonic signal, backscattering noise from grain interface decreases the SNR of received signal. In this paper, SSP(split-spectrum processing) based on the constant FBR decomposition method has been applied to enhance the SNR. This algorithm helps to find optimal parameters of filter bank through a simple theory and has an advantage that reduce the signal processing time compared with the conventional constant bandwidth decomposition method. In this experiment, the 304 stainless steel sample is heat-treated and received ultrasonic signal is processed by SSP using the constand bandwidth decomposition method and the constand FBR decomposition method enhanced the SNR by 1.4 dB and reduced the required number of filters by 4 compared with the constant bandwidth decomposition method.

In this paper, a robust control scheme of a nonlinear system with norm bounded uncertainty is studied. The proposed algorithm is based on variable structure systems (VSS) theory. the sliding mode which is robust to plant uncertainty and disturbances is obtained by regulating a sliding surface equation. This VSS control law can improve the robustness of control systems by adjusting the minimum reaching velocity in a reaching phase. A numerical example is given to verify the effectiveness of the control law.

이산시간 시스템에서 선형분수변환(LPT : linear fractional transformation)으로 표현된 $H^{\infty}$ 제어문제를 체인스케터링표현(CSD : chain scattering description)으로 나타내어 (J,J')-ossless 소인수분해를 이용하여 준최적 $H^{\infty}$ 제어문제를 해결하였다. LFT를 CSD형태로 변환하기 위해서는 표준플랜트의 $P_{21}$의 역행렬이 존재하여야 한다. 본 논문에서는 $P_{21}$의 역행렬이 존재하지 않는 4-블럭문제에서도 LFT를 CSD로 변환하는 방법을 제시하고 이렇게 변환된 행렬을 (J,J')-lossless 소인수분해함으로서 모든 준최적 $H^{\infty}$ 제어기를 매개변수화하였다. 또한 제안한 방법은 단지 두개의 리카티 방정식을 풀므로서 이산시간 시스템의 준최적 $H^{\infty}$ 제어문제를 해결할 수 있음을 보였다.

This paper deals with a new design scheme to reduce the array multiplication time by modifing the structure of the CSA array in the conventional array miltyplier. A circuit with the suggested scheme is designed and simulated. The suggested scheme is to assign the inputs of addend and augend faster than carry input to the CSA. It is shown that the operation time of the CSA array is reduced to 50%.

This paper proposes a divider structure that combines a carry-propagation-free division algorithm using RSD number system and a self-timed ring structure. The self-timed ring structure enables the divider to compute at a speed comparable to that of combinational array dividers with less silicon area. By exploiting the carry-propagation-free division algorithm, we can achieve further reduction of silicon area and computation time. The algorithm and structure of the proposed divider have been successfully verified through VHDL modeling and simulation. Preliminary experimental results show the effectiveness of the algorithm and structure.

Inverse Chebyshev function can realize the same order of Chebyshev function nuder the same specification. In general, inverse Chebyshev function has the preferable characteristics in terms of the delay characteristics and the time-domain performances compare with Chebyshev function. However, for the even order n, inverse Chebyshev function does not realize in the doubly-terminated ladder network which has preferable sensitivity characteristics because of the finite value at ${\omega}={\infty}$. In this paper, the modified inverse Chebyshev function with $\mid$H($j^{\infty}$$\mid$=0 s proposed to realize the passive doubly-terminated ladder network for the n even or odd. The modified inverse Chebyshev function characteristics ars studied in the frequency and time domain, and then, realize the passive doubly-terminated ladder network.

A new efficient contour following algorithm for connected-component labeling processing is proposed. The basic idea of the algorithm is that the total number of downward chain codes is the same as one of upward chain codes along the closed contour. If the chain code direction is upward, then region start mark is assigned at the chain code departure pixel and if the chain code is downward, then region end mark is assigned at the chain code arrival pixel. The proposed algorithm extracts directly the contour information from only the current direction information of chain. This makes the algorithm simple and fast and requires less memory with comparison to the conventional algorithms.The proposed contour following algorithm can be applied to the various kind of image processing such as region filling, restoration and region feature extraction.

In this paper, we propose a new motion vector smoothing scheme which has high reliability and coding efficiency of motion vectors. In motion vector filed, groupings are made using angle and magnitude of the motion vectors. In each group, the unreliable motion vectors are corrected by the motion vector smoothing. In 3$\times$3 window, motion vector of the current block is made to group, and the motion vector smoothing is performed only in that group. Result of computer simulation shows much improvement of the reliability of motion vectors. Moreover, coding bits of the motion vector are diminished from 0.967 to 6.773 bits per block.

Well-trained neural networks have low sensitivity to input errors. Also, the sensitivity to weigh errors must be considered when implementing neural networks with hardware of limited precision. In this paper, we derive the sensitivity of the Madaline to weight perturbation or input errors in terms of the trained weights, the input pattern, and the variance of weight perturbation or the probability of input errors. The result is verified with a simulation of the Madaline recognizing handwritten digits.

This paper presents an output arbitrator for input buffering ATM (Asynchronous Transfer Mode) switches using neural networks. To avoid blocking in ATM switches with blocking characteristics, it is required to buffer ATM cells in input buffer and to schedule them. The N$\times$N request matrix is divided into N/16 submatrices in order to get rid of internal blocking systematically in scheduling phase. The submatrices are grouped into N/4 groups, and the cells in each group are switched alternatively. As the window size of input buffer is increases, the number of input cells switched in a time slot approaches to N. The selection of nonblocking cells to be switched is done by neural network modules. N/4 neural network modules are operated simultaneously. Fast selection can be achieved by massive parallelism of neural networks. The neural networks have 4N neurons and 14N connection. The proposed method is implemented in C language, and the simulation result confirms the feasibility of this method.

In this paper, an algorithm of CBAM(Combination Bidirectional Associative Memory) model proposes, analyzes and tests CBAM model `s performancess by simulating with recalls and recognitions of patterns. In learning-procedure each correlation matrix of training patterns is obtained. As each correlation matrix's some elements correspond to juxtaposition, all correlation matrices are merged into one matrix (Combination Correlation Matrix, CCM). In recall-procedure, CCM is decomposed into a number of correlation matrices by spiliting its elements into the number of elements corresponding to all training patterns. Recalled patterns are obtained by multiplying input pattern with all correlation matrices and selecting a pattern which has the smallest value of energy function. By using a CBAM model, we have some advantages. First, all pattern having less than 20% of noise can be recalled. Second, memory capacity of CBAM model, can be further increased to include English alphabets or patterns. Third, learning time of CBAM model can be reduced greatly because of operation to make CCM.

Both the neural network classifier utilizing multi-layer perceptron and the linear tree classifier composed of hierarchically structured linear discriminating functions can form arbitrarily complex decision boundaries in the feature space and have very similar decision making processes. In this paper, a new method for automatically choosing the number of neurons in the hidden layers and for initalzing the connection weights between the layres and its supporting theory are presented by mapping the sequential structure of the linear tree classifier to the parallel structure of the neural networks having one or two hidden layers. Experimental results on the real data obtained from the military ship images show that this method is effective, and that three exists no siginificant difference in the classification acuracy of both classifiers.

The Max-Min CRI method , a traditional inference method , has three problems: subjective formulation of membership functions, error-prone weighting strategy, and inefficient compositional rule of inference. Because of these problems, there is an insurmountable error region between desired output and inferred output. To overcome these problems, we propose an Intelligent Neuro-Fuzzy System (INFS) based on fuzzy thoery and self-organizing functions of neural networks. INFS makes use of neural networks(Error Back Propagation) to solve the first problem, and NCRI(New Max-Min CRI) method for the second. With a proposed similarity measure, NCRI method is an improved method compared to the traditional Max-Min CRI method. For the last problem, we propose a new defuzzification method which combines only the appropriate rules produced by the rule selection level. Applying INFS to a D.C. series motor, we can conclude that the error region is reduced and NCRI method performs better than Max-Min CRI method.

The binary image normalization method in image processing can be used in several fields, Especially, its high speed processing method and its hardware implmentation is more useful, A normalization process of each character in character recognition requires a lot of processing time. Therefore, the research was done as a part of high speed process of OCR (optical character reader) implementation as a pipeline structure with host computer in hardware to give temporal parallism. For normalization process, general purpose CPU,MC68000, was used to implement it. As a result of experiment, the normalization speed of the hardware is sufficient to implement high speed OCR which the recognition speed is over 140 characters per second.