When vector quantization is used in low rate image coding (e.g., R<0.5), the primary problem is the tremendous computational complexity which is required to search the whole codebook to find the closest codevector to an input vector. Since the number of code vectors in a vector quantizer is given by an exponential function of the dimension. i.e., L=2$^{nR}$ where Rn. To alleviate this problem, a multiple shell structure of hypercube feature maps (MSSHFM) is proposed. A binary HFM of k-dimension is composed of nodes at hypercube vertices and a multiple shell architecture is constructed by surrounding the k-dimensional hfm with a (k+1)-dimensional HFM. Such a multiple shell construction of nodes inherently has a complete tree structure in it and an efficient partial search scheme can be applied with drastically reduced computational complexity, computer simulations of still image coding were conducted and the validity of the proposed method has been verified.

In this paper, we address the issues related to storage and retrieval of continuous media (CM)data we face in designing multimedia on-demand (MOD) storage servers. To support the two orthogonal factors of MOD server design, i.e., storage and retrieval of CM data, this paper discusses the techniques of disk layout, disk striping and real-time disk scheduling, which are integrated as a combined solution to the high- performance MOD storage subsystem. The proposed clustered striping technique enables either a multiple-disk or a parallel system to guarantee a continuous retrieval of CM data at the bandwidth required to support user playback rate by avoiding the formation of I/O bottlenecks.

In TFT-LCDs, as the display size area becomes larger, and the resolution higher, we have to consider the image degradation effects due to the incorporation of the TFT-LCD parameters such as the data-line resistance, the common electrode resistance, the data-line to common parasitic capacitance, and the output characteristics of driver ICs. One of the degradation effects is crosstalk resulting from the coupling between the source bus-line and common electrode. Since a source signal which represents a large number of display data is supposed to vary frequently, the common signal level is affected through the coupling effect, resulting in the degradation of nearby pixel drive signals. Therefore, we proposed a method to compensate for this source-common electrode coupling effect, we also designed and experimented the feasibility of our crosstalk compensation circuit in the actual TFT-LCD. We saw that the newly designed compensation circuit greatly reduced the crosstalk in display pattern image.

In this paper, the method to incorporate TCVQ(Trellis Copded Vector Quantizer) into the encoding of the wavelet trans formed(WT) image followed by a variable length coding(VLC) or an entropy coding(EC) is considered. By WT, an original image is separated into 10 bands with various resolutions and directional components. TCVQ used to compress these WT coefficients is a finite state machine that encodes the input source on the basis of the current input and the current state. Wavelet basis used in this paper is designed by orthogonal spline function. A modified set partitioning algorithm to Wang's is also presented. A simple modification to Wang's algorithm gives a highly time-efficient result. Proposed WT-TCVQ encoder shows a very competitive result, giving 37.46dB in PSNR at 1.002bpp when encoding 512$\times$512 LENA.

For the factory automation(FA) of production or assembly lines, computer vision techniques have been widely used for the recognition and position-control of objects. In this application, it is very important to analyze characteristic features of each object and to find an efficient matching algorithm using the selected features. If the object has regular or homogeneous patterns, the problem is relatively simple. However, If the object is shifted or rotated, and if the depth of the input visual system is not fixed, the problem becomes very complicated. Also, in order to understand and recognize objects with concentric noise patterns, it is more effective to use feature-information represented in polar coordinates than in cartesian coordinates. In this paper, an algorithm for the recognition of objects with concentric circular noise-patterns is proposed. And position-conrtol information is calculated with the matching result. First, a filtering algorithm for eliminating concentric noise patterns is proposed to obtain concentric-feature patterns. Then a shift, rotation and scale invariant alogrithm is proposed for the recognition and position-control of objects uusing invariant feature information. Experimental results indicate the effectiveness of the proposed alogrithm.

An effective algorithm for implementation of unmanned surveillance system which detects moving object from image sequences, predicts the direction of it, and drives the camera in real time is proposed. Outputs of proposed algorithm are coordinates of location of moving object, and they are converted to the values according to camera model. As a pre- processing, extraction of moving object and shape discrimination are performed. Existence of the moving object or scene change is detected by computing the temporal derivatives of consecutive two or more images in a sequence, and this result of derivatives is combined with the edge map from one original gray level image to obtain the position of moving object. Shape discri-mination(Target identification) is performed by analysis of distribution of projection profiles in x and y directions. To reduce the prediction error due to the fact that the motion cha- racteristic of walking man may have an abrupt change of moving direction, an order adaptive lattice structured linear predictor is proposed.

In this paper, we propose a full-search block-matching algorithm with early retirement, which can be applied to a 1-D systolic array of processing elements (PE's) for fast motion estimation. In the proposed algorithm, a PE is retired when its current accumulated sum is equal to or larger than the current minimum MAD. If all PE's are retired, the MAD calculation is stopped for the current array position and is started for the next one in the search window. Simulation results show that the optimum motion vector is always found with less computation, the total computation cycles for motion estimation are decreased to about 60%, and the power dissipation in the PE's is reduced to about 40-60%.

The edge detection process serves to simplify the analysis of images by drastically reducing the amount of data to be processed, while preserving useful structural informations about object boundaries. At first, this paper proposes an edge detection algorithm to reduce the amount of computation. The gradients of pixels are calculated by using first order differential equations on the pixels with even rows and even columns or odd rows and odd columns, and they are compared with a threshold to decide edges. As a result, the computational complexity is reduced to one third or one forth compared with the provious ones. To enhance the accuracy of edge detection, a method with the adaptive threshold for each pixel window which is calculated by using characteristic values is proposed. In this case, the performance can be improved since the threshold is calculated properly for each window according to the local characteristics of corresponding window.

Current test has recently been known to be a promising testing method in CMOS VLSI because conventional voltage test can not make sure of the complete detection of bridging, gate-oxide shorts, stuck-open faults and etc. This paper presents a new BIC(built-in current sensor) for the internal current test in CMOS logic circuit. A single phase clock is used in the BIC to reduce the control circuitry of it and to perform a self- testing for a faulty current. The BIC is designed to detect the faulty current at the end of the clock period, so that it can test the CUT(circuit under test) with much longer critical propagation delay time and larger area than conventional BICs. The circuit is composed of 18 devices and verified by using the SPICE simulator.

In this study, we have developed a prototype of the automatic defects detection system for capsule inspection using the computer vision techniques. The subjects for inspection are empty hard capsules of various sizes which are made of gelatine. To inspect both sides of a capsule, 2-stage recognition is performed. Features we have used are various lengths of a capsule, area, linearity, symmetricity, head curvature and so on. Decision making is performed based on average value which is computed from 20 good capsules in training and permission bounds in factories. Most of time-consuming process for feature extraction is computed by hardware to meet the inspection speed of more than 20 capsules/sec. The main logic for control and arithmetic computation is implemented using EPLD for the sake of easy change of design and reduction in time for developement. As a result of experiment, defects on size or contour of binary images are detected over 95%. Because of dead zone in imaging system, detection ratio of defects on surface, such as bad joint, chip, speck, etc, is lower than the former case. In this case, detection ratio is 50-85%. Defects such as collet pinch and mashed cap/body seldom appear in binary image, and detection ratio is very low. So we have to process the gray-level image directly in partial region.

A new automatic tracking & recognition algorithm which decides the opening & the closing states of subject's eye and isn't affected by the subject's background is proposed. And it was tested in circumstances in which subject's background was not restricted using the developed system, ATRS(Automatic Tracking & Recognition System). The significant characteristic of the ATRS is new movement detection of object that is a body in motion with accelated velocity and it doesn't need any extra hardware except a formal CCD camera and an image grabber but it works so well and so fast. The ATRS would be particularly well suited to a way of communications of patients in a hospital, who can not communicate otherwise.

In this paper, we propose an error diffusion image halftoning technique based on the noise spectrum shaping. The new technique can arbitrarily control the shape of the display error spectrum whereas conventional halftoning algorithms have been known to minimize dc errors only in which case edge information cannot be properly rendered. As a method for estimating the error diffusion coefficients, a least mean square (LMS) approach is adopted.

When the LMS algorithm is employed in the transversal filter structure, the computational complexity can be kept reasonably low. However, if the impulse response to be estimated is very long or signals involved are highly correlated like a speech the convergence speed becomes slow. The lattice filter is an excellent alternative to improve convergence speed since the lattice structure inherently has the orthogonal property among the backward prediction errors, but at the expense of the excessive computational load. If the input signal to be used can be sufficiently well modeled as a .RHO.-th order autoregressive(AR) process, the reflection coefficients after the .RHO.- th stage will be close to zero. Then, instead of employing the full lattice structure, the joint lattice filter structure can be implemented in conjunction with the transversal filter structure after the .RHO.-th stage. We propose, in this paper, this new lattice/transversal joint structure, and we will call it the reduced lattice filter. Using the reduced lattice filter, we are now able to achieve the performance as good as that of the lattice filter, while maintaining the complexity as low as that of the transversal filter. The proposed filter is particularly useful for an acoustic echo canceller due to the highly correlatedness nature of speeches and the long and frequently changing echo paths.

Due to the low loss, broadband and accurate short time delay properties of optical fiber, it has attracted as a delay medium for high speed and broad-band signal processing. In this paper, we consider the coherent optical fiber filter of IIR lattice structure, which uses coherent light sources and consists of directional couplers whose coupling coefficients are restricted between 0 and 1. Considering restrictions of directional coupler, the design formulae and condition for realibility of optical fiber filter of IIR lattice structure which makes the optimal use of optical signal energy are derived.

In this paper, we propose recurrent neural prediction models (RNPM), recurrent neural networks trained as a nonlinear predictor of speech, as a new connectionist model for speech recognition. RNPM modulates its mapping effectively by internal representation, and it requires no time alignment algorithm. Therefore, computational load at the recognition stage is reduced substantially compared with the well known predictive neural networks (PNN), and the size of the required memory is much smaller. And, RNPM does not suffer from the problem of deciding the time varying target function. In the speaker dependent and independent speech recognition experiments under the various conditions, the proposed model was comparable in recognition performance to the PNN, while retaining the above merits that PNN doesn't have.

Maps are one of the most complicated types of drawings. Drawing recognition technology is not yet sophisticated enough for automated map reading. To automatically extract a road map dircetly from complicated topographical maps, a very sophisticated algorithm is needed, since the image generally involvfes such complicated patterns as symbols, characters, residential sections, rivers, railroads, etc. This paper proposes a new feature extraction method based on the morphology. We apply this method to extract complete set of road segments from topographical maps. The proposed method successfully extract road segments from various areas.

In this paper, we propose the filtered-x least mean fourth (FXLMF) algorithm where the error raised to the power of four is minimized and analyze its convergence behavior for a multiple sinusoidal acoustic noise and Gaussian measurement noise. Application of the FXLMF adaptive filter to active noise control requires to estimate the transfer characteristics of the acoustic path between the output and the error signal of the adaptive controller. The results of the convergence analysis of the FXLMF algorithm indicate that the effects of the parameter estimation inaccuracy on the convergence behavior of the algorithm are characterized by two distinct components : Phase estimation error and estimated gain. In particular, the convergence is shown to be strongly affected by the accuracy of the phase response estimate. Also, we newly show that the convergence behavior can differ depending on the relative sizes of the Gaussian noise and the convergence constant.

This paper introduces an isolated word recognition system realized on a neurocomputer called E-MIND II, which is a 2-D torus wavefront array processor consisting of 256 DNP IIs. The DNP II is an all digital VLSI unit processor for the EMIND II featuring the emulation capability of more than thousands of neurons, the 40 MHz clock speed, and the on-chip learning. Built by these PEs in 2-D toroidal mesh architecture, the E- MIND II can be accelerated over 2 Gcps computation speed. In this light, the advantages of the E-MIND II in its capability of computing speed, scalability, computer interface, and learning are especially suitable for real time application such as speech recognition. We show how to map a TDNN structure on this array and how to code the learning and recognition algorithms for a user independent isolated word recognition. Through hardware simulation, we show that recognition rate of this system is about 97% for 30 command words for a robot control.

We propose a design scheme of a binary decision tree and apply it to the tire tread pattern recognition problem. In this scheme, a binary decision tree is constructed by using fuzzy C-means( FCM ) algorithm. All the available features are used while clustering. At each node, the best feature or feature subset among these available features is selected based on proposed similarity measure. The decision tree can be used for the classification of unknown patterns. The proposed design scheme is applied to the tire tread pattern recognition problem. The design procedure including feature extraction is described. Experimental results are given to show the usefulness of this scheme.