Search

Assuming the leakage of hazardous liquid chemicals in factories and chemical plants caused by natural disasters or human error, it is important to precisely estimate exposure risk and design appropriate evacuation plans and risk‐reduction measures. The exposure risk of indoor residents to leaked chemicals may be estimated using direct measurements or theoretical/numerical estimations of gas‐phase concentrations. Although mathematical models of gas‐phase concentration in the workplace are essential for predicting exposure concentrations at the design stage, practical models have not yet been fully established. We previously proposed practical numerical models to assess indoor gas‐phase concentration distributions after the accidental leakage of liquid chemicals using liquid toluene as a representative hazardous chemical. In this study, the leakage of liquid ethyl acetate, a commonly used organic solvent, into mortar flooring was reproduced in a small test chamber, and numerical analysis was performed. The results showed that the time history of the chamber exhaust concentration had a sharp peak and differed significantly from that of toluene. Three‐dimensional computational fluid dynamics (CFD) analysis incorporating practical ethyl acetate emission models reproduced the external and internal emission characteristics of leaked ethyl acetate on the mortar material. The models showed reasonable agreement with the experimental data. [ABSTRACT FROM AUTHOR]

This paper proposes a novel and general method of glare generation based on wave optics. A glare image is regarded as a result of Fraunhofer diffraction, which is equivalent to a 2D Fourier transform of the image of given apertures or obstacles. In conventional methods, the shapes of glare images are categorized according to their source apertures, such as pupils and eyelashes and their basic shapes (e.g. halos, coronas, or radial streaks) are manually generated as templates, mainly based on statistical observation. Realistic variations of these basic shapes often depend on the use of random numbers. Our proposed method computes glare images fully automatically from aperture images and can be applied universally to all kinds of apertures, including camera diaphragms. It can handle dynamic changes in the position of the aperture relative to the light source, which enables subtle movement or rotation of glare streaks. Spectra can also be simulated in the glare, since the intensity of diffraction depends on the wavelength of light. The resulting glare image is superimposed onto a given computer-generated image containing high-intensity light sources or reflections, aligning the center of the glare image to the high-intensity areas. Our method is implemented as a multipass rendering software. By precomputing the dynamic glare image set and putting it into texture memory, the software runs at an interactive rate. [ABSTRACT FROM AUTHOR]

Visual information in 3D space has become the subject of much research in fields like 3D image communications and virtual reality. The huge amount of visual information in 3D space, however, makes the establishment of an efficient compression technique indispensable. In this regard, the authors have already proposed the concept of “space coding” that makes use of the redundancy in space itself, and have been investigating a specific coding technique that uses a multidimensional discrete cosine transform (DCT). In this paper, we extend this technique and propose a new coding technique that takes the structural characteristics of the object into account. To be more specific, we propose the concept of a “virtual object surface” as a technique for describing an object in structural terms taking compression rate into account, and attempt to improve coding efficiency with this technique. We also describe experiments performed to compare this technique with other techniques, and demonstrate the feasibility of the proposed technique on the basis of the results obtained. In addition, we examine a coding technique for performing high-speed image synthesis from compressed data. © 2001 Scripta Technica, Syst Comp Jpn, 32(12): 47–59, 2001 [ABSTRACT FROM AUTHOR]

In the field of 3D visual communications, a method of description of 3D images that would be neutral (independent) with respect to specific techniques for 3D input–output is desirable. In this context, the authors have proposed the concept of “ray-space” as a ray-based description method for 3D image data. In this paper, the concept of “ray-space” is employed to offer a new method for simultaneous description of 3D objects and space, as well as a technique for interactive manipulation in virtual space built using real images. Synthesis of images as seen from an arbitrary viewpoint in real space means extraction of ray data on a curved surface described by some basic formula in ray-space, using information about the viewpoint location. In this study, real-time processing is pursued; for this purpose, a reference table is employed for efficient extraction of ray data. Experiments were carried out concerning virtual manipulation of real objects in near-real time by means of overwriting ray-space data with ray data of the 3D object under consideration. The experiments confirmed that the proposed method offered efficient description of 3D space data and object manipulation. © 1999 Scripta Technica, Syst Comp Jpn, 30(14): 12–22, 1999 [ABSTRACT FROM AUTHOR]

This article relates to interactive command coding with active participation by the user in the coding process through his interactive operations, and structuralization of a moving image according to its substance. This coding process enables a low rate transmission with the preferential transmission of the substance of the original image, which is performed by using the user's interactive operations for extracting characteristics of objects relating to their shapes, texture, and motions, and by appropriately assigning a detailed description-based or simplified description-based technique, depending on the substance of the original image. In addition, it enables easy browsing for capturing outlines on the decoding side, and easy application even to nonlinear image editing, by modifying the motion-related parameters and also modifying the feature description technique to another technique. Concrete examples including test examples are given to explain this approach. © 1999 Scripta Technica, Syst Comp Jpn, 30(13): 73–83, 1999 [ABSTRACT FROM AUTHOR]

A new method of designing control sensors for intelligent transport systems (ITS) is proposed in which blooming is estimated by partial autocorrelation analysis, the distance from the focal point to the object is derived, and the distance to the object is measured. In partial autocorrelation analysis, forward prediction and backward prediction are carried out simultaneously so that the point spread function convoluted with the blooming image is linearly predicted and the amount of blooming is estimated. The two-dimensional analysis requires a significant amount of processing and highly accurate operation. Therefore, in this method, by preprocessing in the frequency dimension, only the horizontal components are extracted for analysis. From the experimental results, it is found that distance estimation from a single image is possible near the focal length (2.5 m–1.7 m). © 1999 Scripta Technica, Electron Comm Jpn Pt 2, 82(1): 23–33, 1999 [ABSTRACT FROM AUTHOR]

The authors investigated construction of an extended-order statistic filter to remove non-Gaussian white noise added to image signal. First, they devised a differential-order statistic filter using signal-differential ordering instead of a signal-value ordering and estimated its noise reduction performance. Next, they proposed an extended-order statistic filter which unifies the three-order information of time, signal value, and signal differences obtained from a signal within the filter window. Preparing the 3-dimensionalized filter coefficient space and using filter coefficients chosen sequentially in the three different orders among them allow effective noise reduction performance to be obtained. In addition, as a method for resolving a problem on the design method of filter coefficients, investigation was made using a synthesized training sequence. Extracting parameters necessary to a linear AR model from an objective signal, it is shown that the filter design using a training sequence synthesized from it gives noise reduction performance similar to the one at the optimal design. [ABSTRACT FROM AUTHOR]

This paper proposes the adaptive discrete cosine transform coding with vector quantization (adaptive DCT-VQ), as a highly efficient coding for the picture signal. In the proposed method, the vector quantization is performed by applying the discrete cosine transform to the picture subblock and partitioning the transformed domain into several vectors. By regarding the transformed domain as more than one vector, the complexity in the coding is reduced. Each vector quantizer is designed according to the standard probability distribution of the cosine transform coefficient, and the universality of the vector quantizer is realized. The mode-selecting type adaptive control is introduced, to effect the adaptive control of the coding parameters according to the local property of the picture. To cope with the large bit allocation resulting from the adaptive control, each vector quantizer has two-stage structure thereby reducing the coding complexity. As a result of the performance evaluation by simulation, it was verified that the proposed coding method has a higher performance than the conventional adaptive cosine transform coding and has much less complexity than the vector quantization in the original space domain. [ABSTRACT FROM AUTHOR]

A digital filter is developed to estimate signals containing discontinuities from observations contaminated with low-amplitude white noise. Called an &epsilon-Kalman filter, this is a Kalman filter combined with a nonlinear function. It is very effective for signals with known statistical properties. Principles of operation and realization of the filter are discussed. It is compared with another nonlinear filter called an ε-filter also proposed by the authors. Results of computer simulations are shown to demonstrate the effectiveness of the new filter. Pictorial examples are also provided. [ABSTRACT FROM AUTHOR]

This paper presents a new design method for a frequency sampling adaptive equalizer using the least-squares method and reports the results obtained in analyzing the equalizer performance. This system employs an adaptive equalizer configuration in which a cyclic code equal to the tap length is used as a test signal with the received signal and the reference signal each undergoing a discrete Fourier transformation using a frequency sampling filter and applying the least-squares method in a one-to-one relationship to each DFT coefficient. A condition for the implementation of this system is that the test signal be a cyclic code, but this system requires fewer arithmetic operations than any other system employing the least-squares method. The system presented here employs a high-speed convergence algorithm in the initial training period and after convergence is accomplished it switches to a gradient method using a frequency sampling adaptive equalizer. In simulation tests with telephone lines it was found that this equalizer converged in about one-third the time required for the transversal adaptive equalizer. [ABSTRACT FROM AUTHOR]

In a previous paper we presented procedures of analyzing causality between a time series based on information theory and information flow between time series based on a concept of directed transinformation, which is an information quantity with a direction. In principle, the directed transinformation is defined based on joint probability density between time series. As the length of the time series becomes larger, however, the volume of computation needed also increased exponentially. In this paper, we will discuss efficient computation algorithms. First, we assume Gaussian properties with the time series to present the "correlation function method" as a means of computing of the directed transinformation from correlation function. Second, we present a "linear production model method" in which the information quantity is computed by means of impulse response of a linear production model of the time series. Also, we proved that the linear production model can easily be composed by auto-regression if the time series is steady. Finally, we discussed the validity of time-series analysis based on the directed transinformation through analysis of an artificially produced simulated series. We have concluded that the proposed computation method can provide correct analysis even if the causality of the time series can not be analyzed correctly based on the correlation function. [ABSTRACT FROM AUTHOR]

This paper presents a method for constructing a cyber-space with a high degree of reality by integrating the generation of images with an arbitrary viewpoint from the actually taken image, with the computer graphics (CG) techniques based on the geometric model. There are intensive studies of image-based rendering, which reproduces space based on actually taken images. This paper presents a new method that combines such a method and an existing geometric model, while maintaining the features of those techniques. In this paper, the ray space is used for the image-based rendering, and efforts are made to define the base plane so that the entire surroundings of the object can be observed. In order to represent the space with the mixed ray-space data and the geometric model, the same data structure as used in VRML is employed. © 1998 Scripta Technica, Syst Comp Jpn, 29(9): 21–31, 1998 [ABSTRACT FROM AUTHOR]

This paper proposes a new coding method using structural features such as shape and motion in a moving image so that they are described as plane motion in a 3D space. In this method, a layered representation (a moving image is described on a 2D projection plane) is extended by using a setting representation so as to handle 3D data. In the setting representation, 3D parameters which describe the orientation of a plane and the motion of a rigid body are needed for the regions that move coherently on the projection plane. Parameters can be estimated stably by applying a conventional 3D parameter estimation algorithm (which uses the square-error criterion based on the assumption of equal intensity between adjacent frames) to multiple frames. This paper also proposes a method of integration of the rigid-body motion of multiple regions so that the motion of images can be described more efficiently. The proposed method is compared with the layered representation using 2D data, and a conventional waveform-based method. The results show the effectiveness of the proposed method. © 1998 Scripta Technica, Syst Comp Jpn, 28(10): 50–59, 1997 [ABSTRACT FROM AUTHOR]

The human face includes a great deal of information related to individual features. If it were possible to describe facial images by means of parameters, one could think of a variety of applications. In this paper, wireframe models are applied to facial images, and orthogonal facial bases are found using principal-component analysis, while shape-related data and texture-related data are treated separately. By means of these facial bases, effective parametric description of facial images is achieved. The proposed method supports not only analysis of facial images but also image synthesis using obtained parameters. © 1998 Scripta Technica, Syst Comp Jpn, 28(10): 77–83, 1997 [ABSTRACT FROM AUTHOR]

To realize a user-friendly interface where a human and a computer can engage in face-to-face communication, the computer must be able to recognize the emotional state of the human by facial expressions, then synthesize and display a reasonable facial image in response. To describe this analysis and synthesis of facial expressions easily, the computer itself should have some kind of emotion model. By using a five-layered neural network which has generalization and superior nonlinear mapping performance, identity mapping training was performed using parameterized facial expressions. With respect to the space built in the middle layer of the five-layered neural network as an emotion model, emotion space was constructed. Based on this emotion space an attempt was made to build a system which can realize mapping from an expression to an emotion and from an emotion to an expression simultaneously. Moreover, to recognize a facial expression from the actual facial image of a human, the method of extracting the facial parameter from the facial points movement, is investigated. [ABSTRACT FROM AUTHOR]

This paper proposes a method which analyzes an expression from a facial image using a three-dimensional facial model and the extracts the facial expression. First, the head motion the action (such as those of eyebrows, eyes, and lips) are separated from the facial image. This is realized by estimating the three-dimensional facial and by compensating the motion. Next, the expression information is extracted from the separated facial actions in two ways. One is the method to extract successively the facial expressions considering the characteristics of the facial action actions based on the facial muscles. The other is the method to estimated the facial expression as a using the least-square method and regarding the facial actions by the facial muscles as a vector. Those methods are it possible to reconstruct the original expression from the extracted facial expression parameters. Finally, the result of the analysis of the facial expression from the actual image analysis of the facial result of evaluation by a psychologist to demonstrate the usefulness of the proposed method. The image reconstructed from the result of analysis also is compared with the original image. [ABSTRACT FROM AUTHOR]

If an object is a rigid body, it is possible to estimate its three-dimensional structure and motion mathematically from motion information. However, the assumption that objects are rigid bodies includes a disadvantage that it is not stable against noise and that the circumstance of application is restricted. When considering application to natural images or modeling of vision, a general-purpose method for estimating three-dimensional structure and motion which allows discrepancy from rigid bodies is desired. Thus, in this piper, an algorithm is presented for estimating three- dimensional structure and motion incrementally from a Long image sequence using the condition for "smoothness of motion" instead of the rigid body assumption. That is, as a constraint, a physical property is used whereby an object generally performs smooth motion due to inertia and elasticity. Formulating this condition, i.e., smoothness of motion as a stabilizing function for regularization, the three-dimensional information is estimated incrementally so that it is suitable to an input image and motion is motion. By considering smoothness of motion, three-dimensional information of each frame can be stored incrementally, and thus a flexible algorithm that is robust and also applies to the motion of a nonrigid body is obtained. Furthermore, by presenting the results when applying this method to the motion of rigid and nonrigid bodies, the robustness of this method against noise and flexibility is shown which also applies to nonrigid bodies. [ABSTRACT FROM AUTHOR]

Model-based image coding has been highlighted recently as a high-efficiency coding method for TV telephone and TV conference systems. In a model-based coding system, an ultralow-rate image transmission is realized by obtaining common models of a facial image at both sides of a communication and by transmitting only modification parameters between them. However, it is difficult to realize a realtime processing of a model-based coding with a conventional iterative-processing type computer since the amount of material to analyze and synthesize at both sides is very large. Realtime processing is absolutely necessary to realize a practical system using this method, i.e., high-speed processings at both the transmitter and receiver are required. This paper describes a realtime facial image synthesis method based on a multiprocessor construction with a transputer. The transputer is a microcomputer having a communication capability with multiple processors and 10 MIPS CPU performance. Using this system in a pipeline processing with 20 processors, it is possible to synthesize realtime facial images at a rate of about 16 frames per second. If only a part around the lips is transmitted, it is possible to synthesize an image at a rate of about 40 frames per second with a network using five processors. [ABSTRACT FROM AUTHOR]

Video communications play an important role in an information-oriented society. A cost-effective image coding technique capable of rendering high quality video communication services is an important component of any information system. This paper describes an algorithm for image sequence coding using three-dimension- al (3-D) structure and motion information. Each image is treated as a 2-D projection of a 3-D object, with 3-D information derived and used in the coding scheme. Since the coding scheme considers physical 3-D in- formation, it provides data compression at the transmitter and improved image quality at the receiver. Furthermore, because the image is represented in a structure-oriented fashion, it lends itself to image recognition and image understanding. This aids in the realization of high quality communication systems. The proposed algorithm makes extensive use of 3-D structure and motion information in conjunction with 3-D motion interpolation and 3-D motion compensation predictive coding technique to achieve efficient, high quality image sequence coding. A method is described that has 3-D structure and motion information extraction, and an incremental scheme is presented for stable estimation of 3-D structure and motion information of rigid as well as non-rigid bodies. [ABSTRACT FROM AUTHOR]

A facial motion image synthesis method for intelligent man-machine interface is examined. Here, the intelligent man-machine interface is a kind of friendly man-machine interface with voices and pictures in which human faces appear on a screen and answer questions, compared to the currently existing user interfaces which primarily uses letters. Thus what appears on the screen is human faces, and if speech mannerisms and facial expressions are natural, then the interactions with the machine are similar to those with actual human beings. To implement such an intelligent man-machine interface it is necessary to synthesize natural facial expressions on the screen. This paper investigates a method to synthesize facial motion images based on given information on text and emotion, The proposed method utilizes the analysis-synthesis image coding method. It constructs facial images by assigning intensity data to the parameters of a 3-dimensional (3-D) model matching the person in question. Moreover, it synthesizes facial expressions by modifying the 3-D model according to the predetermined set of rules based on the input phonemes and emotion, and also synthesizes reasonably natural facial images. [ABSTRACT FROM AUTHOR]

Generally speaking, it is necessary to extract knowledge from an expert in a given discipline and implement this knowledge into a system when constructing an expert system. However, it is not easy to extract knowledge in such fields as medical diagnosis or pattern recognition because the inference logic depends on the experience and intuition of the expert. This paper proposes an automatic rule extraction mechanism using statistical analysis. In this system, production rules are expressed in the form of the threshold function. Because the threshold function can describe any kinds of inference logic, it is expressed easily as a linear combination of input vectors and weighting coefficients. Thus weighting coefficients can be calculated by the same method as a discriminant function. If only one threshold is defined, general Boolean logic an be expressed. Moreover, an ambiguous inference rule can be expressed when the threshold levels are multidefined and a membership function is defined at each category. Further, the Fuzzy Tree Search algorithm which combines ambiguous inference and tree search is proposed at the end of this paper. This algorithm can search and determine an optimum cluster with little calculation and good performance. In practice, a medical diagnostic system applied to psychiatry which has most ambiguous diagnostic logic, has been constructed based on this algorithm and inference rules have been extracted automatically. By this experiment, Fuzzy Tree Search is as fast as the tree search technique and has as good a performance as a full search clustering technique. [ABSTRACT FROM AUTHOR]

For high-efficiency image compression, previously, an SVD (singular value decomposition)-based coder was developed using vector quantization, called SVD-VQ. This paper proposes an improved quantization SVD-VQ scheme. For every input subblock, the SVD- VQ coder scalar-quantizes a singular value and vector-quantizes two singular vectors, separately. The SVD-VQ decoder reproduces a subblock as the product of these quantization outputs, but does not necessarily produce a reconstruction with the minimum distortion in an image space. This paper develops a quantization scheme where the minimum-distortion reconstruction is always provided in the original image space and presents its design algorithm. The improved SVD-VQ shows A/N performance improvement of 0.5 - 1.0 dB over the conventional SVD-VQ, and is similar in performance to the adaptive DCT (discrete cosine transform) coder. [ABSTRACT FROM AUTHOR]

The authors have previously proposed a quantization scheme in which the ac discrete cosine transform coefficients after normalization are quantized by the gain/shape vector quantizer (called DCT-G/SVQ). It is shown from the relation between the quantization complexity and the S/N ratio that DCT-G/SVQ exhibits a better performance than the discrete cosine transform coding with vector quantization (called DCT-VQ) discussed in the literature [6]. This paper proposes the adaptive DCT-G/SVQ, in which the adaptive control is introduced £or the coding parameters of DCT-G/SVQ. Two methods of adaptive control are considered--the mode- selection and the recursive adaptation. The adaptive DCT-G/SVQ has a better performance than the traditional adaptive discrete cosine transform coding using the scalar quantizer. In other words, the S/N ratio is improved by 1.0 to 1.5 dB in the mode-selection adaptive control coding at the same coding rate. The S/N ratio in the recursive adaptive control coding is improved by 0.5 to 1.0 dB. The adaptive DCT-G/SVQ based on the mode-selection control exhibits nearly the same coding performance as the adaptive DCT-VQ of [9] with the same adaptive control scheme, but is superior since it requires less complexity in quantization. [ABSTRACT FROM AUTHOR]

To realize fast search at vector quantization applied to high-efficiency coding of signals, we propose a partial distance algorithm and a nearest-neighbor search algorithm, both in the orthogonal transform domain. These algorithms provide efficient search for signals such as, for example, picture signals which have a concentration in the transform domain power distribution. Further, we propose a nearest-neighbor search algorithm in orthogonal transform domain with component selection which is effective also for memoryless random information sources with Gaussian distribution. The proposed algorithm is applied to vector quantizers of picture signals, separating-mean-normalized vector quantizers, and vector quantizers of memoryless Gaussian sources and its efficiency is demonstrated for these cases. Particularly, in the case of vector quantizers of picture signals, it is shown that the proposed method requires fewer multiplications to determine the optimum reconstructing vector than the conventional tree search algorithm. [ABSTRACT FROM AUTHOR]

A gain/shape vector quantizer for orthogonal transform domain has been designed based on a previous paper [11] on a gain/shape vector quantizer for a spherically symmetric random source. Such a design method is used in the design of a discrete cosine transform encoder with gain/shape vector quantizer. This encoder is similar to that described in [7] except in the following: (i) it has a gain/shape codebook; and (ii) its normalized coefficients of discrete cosine transform are modeled as having spherically symmetric distribution. The S/N rendition of this encoder is 1.0 - 2.5 dB better than that of the DCT coding scheme with Max-type scalar quantizer at the same bit rate. Furthermore, in spite of its similarity to the conventional discrete cosine transform coding scheme with vector quantizer [7], the encoder introduced in this paper delivers higher S/N using the same number of multiplications for the quantization of normalized coefficients, and/or the same amount of memory for the codebook. [ABSTRACT FROM AUTHOR]

The probability density of normalized ac orthogonal transform coefficients of image data can be modeled as a spherically symmetric distribution. On the other hand, the gain/shape vector quantizer [8] proposed by Buzo et al. is expected to provide a good quantization performance and a flexible design method for spherically symmetric random source. Therefore, it may be possible to realize a flexible vector quantizer by designing a gain/shape vector quantizer in the orthogonal transform domain. With this expectation the asymptotic performance and optimal allocations of the number of quantization levels of the gain/shape vector quantizer for an n-dimensional spherically symmetric random source was derived and it was shown that for n ≤ 10 the gain/shape vector quantizer yields the asymptotic performance between its upper bounds derived y Zador's [10] and those by Conway and Sloane [11], and that for n > 11 its asymptotic performance is similar to the Zador's upper bounds. Also, comparison with the asymptotic performance of scalar quantizers [12] was made and it was shown that possible improvement of coding performance can be obtained by applying the gain/shape vector quantizer to orthogonal transform image conding, for which conventionally scalar quantizers have been used. Furthermore, a simple and flexible design method for the suboptimal gain/shape vector quantizer for a given spherically symmetric distribution was presented. [ABSTRACT FROM AUTHOR]

In a picture data base the picture quality required depends on the actual application of the picture signal. Therefore efficient reversible coding systems are needed. On the other hand, users demand easy searching of the required pictures. Considering the above facts, two kinds of information preserving transmission systems-System 1 and System 2-are proposed for picture data base treating 8-bit PCM picture signals. In these systems, first, reference picture signal for user's reference are transmitted. Next, a supplementary transmission is performed. In System 1 a degenerate differential signal between PCM picture signal and reference picture signal is encoded and transmitted. In system 2, PCM picture signal is reversibly encoded and transmitted directly. In this paper, four kinds of coding systems are considered as coding system for reference pictures. The best performance has been achieved by discrete cosine transform pictures. The best performance has been achieved by discrete cosine transform coding. In addition, this paper presents a method for determining the better system in performance of supplementary transmission of the two systems on the basis of signal-to-noise ratio of the reference picture. Finally, the realization of a transmission systems is shown which combines the advantages of both Systems 1 and 2 by choosing the better system on the basis of coding performance at each reference picture. [ABSTRACT FROM AUTHOR]

The impression of the human face is very important in daily life. However few studies of it have been made from the technical point of view. This is because impressions of a face are formed in the mind of a person and it is not easy to handle this computationally. This paper proposes a method to synthesize an average face of a group of individuals in order to analyze the “impression” quantitatively. A facial space can be constructed by using multiple average faces. This facial space assists in discriminating among several impressions in changing the impression of the original faces. © 1998 Scripta Technica, Electron Comm Jpn Pt 3, 81(11): 29–35, 1998 [ABSTRACT FROM AUTHOR]

This paper presents a new analysis method for the nonstationary phenomenon using modeling. Method efficacy is investigated and effectiveness is demonstrated. Nonstationarity is modeled into the kinetics of the generating process. The behavior of the nonstationary process is represented by the characteristics at a point on the time axis (instantaneous characteristic) and the characteristics of its time variation (time-variant characteristics). By estimating the nonstationary spectrum, the instantaneous characteristics are determined and then the time-variant characteristics are analyzed by applying the kinetic model. The nonstationary phenomenon is first modeled using kinetic energy. Characteristics at a point on the time axis are regarded as a point of mass. Position, velocity, and acceleration are determined; these parameters signify a model with time-varying characteristics. The arising time of the event is determined, based on the relations among the parameters. The characteristic plane is constructed using these parameters, and nonstationarity is evaluated on the plane. The proposed algorithm is then evaluated using an artificial signal and the model is validated using a speech signal. This analysis yields a new finding that deals with the characteristics of the phoneme, thus dem- onstrating the usefulness of the proposed method. [ABSTRACT FROM AUTHOR]

The automatic synthesis of the facial expression by computer is one of the most basic techniques to be applied in various fields. This paper describes a method to synthesize various facial expressions by modification of the 3-D model of the face. The 3-D facial model is obtained by adjusting the prepared 3-D facial shape model to the frontal view of the object neutral face image, and by projecting the gray-level informations. This paper proposes first a hierarchical model of facial description units for the hierarchical description of the facial expressions. The action unit (AU) of the FACS (facial action coding system), which is known as a systematic method of describing the facial expression, is considered as a layer in the proposed hierarchical model. AU of FACS is realized on a computer, and the automatic synthesis of the expression is executed. This procedure is based on the muscular structure of the face and allows cancellation of individual differences. At present, 34 kinds of AUs are implemented. The gray-level information for wrinkles and teeth, which are important in the synthesis of expressions, is not contained in the 3-D facial model obtained from the neutral expression. Thus, ouch information is extracted from other face images, and registered as auxiliary information of the 3-D facial model. This leads to synthesis of natural expressions. Finally, numerous examples for the facial expression synthesis are presented to show the usefulness of the method. [ABSTRACT FROM AUTHOR]