In this paper, we study three-dimensional motion estimation using optical flow. We construct a weighted quotient-form objective function that provides an unbiased estimator. Using this objective function with a certain projection operator as a weight drastically reduces the computational cost for estimation compared with using the maximum likelihood estimator. To reduce the variance of the estimator, we examine the weight, and we show by theoretical evaluations and simulations that, with an appropriate projection function, and when the noise variance is not too small, this objective function provides an estimator whose variance is smaller than that of the maximum likelihood estimator. The use of this projection is based on the knowledge that the depth function has a positive value (i. e., the object is in front of the camera) and that it is generally smooth.

Recently applications of Hilbert curves are studied in the area of image processing, image compression, computer hologram, etc. We have proposed a fast Hilbert scanning algorithm using lookup tables in N dimensional space. However, this scan is different from the one of previously proposed scanning algorithms. Making the lookup tables is a problem for the generation of several Hilbert scans. In this note, we describe a method of making lookup tables from a given Hilbert scan which is obtained by other scanning methods.

Understanding unknown objects in images is one of the most important fields of the computer vision. We are confronted with the problem of dealing with the ambiguity of the image information about unknown objects in the scene. The purpose of this paper is to propose a new object recognition method based on the fuzzy relation system and the fuzzy integral. In order to deal with the ambiguity of the image information, we apply the fuzzy theory to object recognition subjects. Firstly, we define the degree of similarity based on the fuzzy relation system among input images and object models. In the next, to avoid the uncertainty of relations between the input image and the 2-D aspects of models, we integrate the degree of similarity obtained from several input images by the fuzzy integral. This proposing method makes it possible to recognize the unknown objects correctly under the ambiguity of the image information. And the validity of our method is confirmed by the experiments with six kinds of chairs.

This paper introduces a new image representation method that is named the projective image representation (PIR). We consider an image as a collage of symmetric segments each of which can be well represented by its projection data of a single orientation. A quadtree-based method is adopted to decompose an image into variable sized segments according to the complexity within it. Also, we deal with the application of the PIR to the image compression and propose an efficient algorithm, the quadtree-structured projection vector quantization (QTPVQ) which combines the PIR with the VQ. As the VQ is carried out on the projection data instead of the pixel intensities of the segment, the QTPVQ successfully overcomes the drawbacks of the conventional VQ algorithms such as the blocking artifact and the difficulty in manipulating the large dimension. Above all, the QTPVQ improves the subjective quality greatly, especially at low bit rate, which makes it applicable to low bit rate image coding.

Computer viruses, hackers, intrusions and ther computer crimes have recently become a serious security problem in information systems. Digital signatures are useful to defend against these threats, especially against computer viruses. This is because a modification of a file can be detected by checking the consistency of the originai file with its accompanying digital signature. But an executable program might have been infected with the viruses before the signature was created. In this case, the infection cannot be detected by signature verification and the origin of the infection cannot be specified either. In this paper, we propose a signature scheme in which one can sign right after the creation of an executable program. That is, when a user compiles a source program, the compiler automatically creates both the executable program and its signature. Thus viruses cannot infect the executable programs without detection. Moreover, we can specify the creator of contaminated executable programs. In our signature scheme, a signature is created from a set of secret integers stored in a compiler, which is calculated from a compiler-maker's secret key. Each compiler is possessed by only one user and it is used only when a secret value is fed into it. In this way a signature of an executable program and the compiler-owner are linked to each other. Despite these measures, an executable program could run abnormally because of an infection in prepro-cessing step, e.g. an infection of library files or included files. An infection of these files is detected by ordinary digital signatures. The proposed signature scheme together with digital signature against infection in the preprocessing step enables us to specify the origin of the infection. The name of the signature creator is not necessary for detecting an infection. So, an owner's public value is not searched in our scheme, and only a public value of a compiler-maker is required for signature verification. Furthermore, no one can use a compiler owned by another to create a proper signature.

GDL is the language whose membership problerm is polynomial-time Turing equivalent to the discrete logarithm problem for a general finite group G. This paper gives a characterization of GDL from the viewpoint of computational complexity theory. It is shown that GDL NP co-AM, assuming that G is in NP co-NP, and that the group law operation of G can be executed in polynomial time of the element size. Furthermore, as a natural probabilistic extension, the complexity of GDL is investigated under the assumption that the group law operation is executed in an expected polynomial time of the element size. In this case, it is shown that GDL MA co-AM if G MA co-MA. As a consequence, we show that GDL is not NP-complete unless the polynomial time hierarchy collapses to the second level.

An invertible length preserving transducer is called a weakly invertible finite automaton (WIFA for short). If the first letter of any input string of length τ + 1 is uniquely determined by the corresponding output string by a WIFA and its initial state, it is called a WIFA with delay τ. The composition of two WIFAs is the natural concatenation of them. The composition is also a WIFA whose delay is less than or equal to the sum of the delays of the two WIFAs. In this paper we derive various results on a decomposition of a WIFA into WIFAs with smaller delays. The motivation of this subject is from theoretical interests as well as an application to cryptosystems. In order to capture the essence of the decomposability problem, we concentrate on WIFAs such that their input alphabets and their output alphabets are identical. A WIFA with size n of the input and output alphabet is denoted by an n-WIFA. We prove that for any n > 1, there exists an n-WIFA with delay 2 which cannot be decomposed into two n-WIFAs with delay 1. A one-element logic memory cell is a special WIFA with delay 1, and it is called a delay unit. We show that for any prime number p, every strongly connected p-WIFA with delay 1 can be decomposed into a WIFA with delay 0 and a delay unit, and that any 2-WIFA can be decomposed into a WIFA wiht delay 0 and a sequence of k delay units if and only if every state of the 2-WIFA has delay k.

Semi-Orthogonally Associative Memory neural network model (SAM) uses the orthogonal vectors in Un = {-1, 1}n as its characteristic patterns. It is necessary to select the optimum characteristic parameter n so as to increase the efficiency of this model used. This paper investigates the dynamic behavior and error correcting capability of SAM by statistical neurodynamics, and demonstrates that there exists a convergence criterion in tis recalling processes. And then, making use of these results, its optimum characteristic parameter is deduced. It is proved that, in the statistical sense, its recalling outputs converge to the desired pattern when the initial similar probability is larger than the convergence criterion and not true otherwise. For a SAM with N neurons, when its characteristic parameter is optimum, its memory capacity is N/2 ln ln N, the information storage capacity per connection weight is larger than 9/23 (bits/weight) and the radius of attractive basin of non-spurious stable state is about 0.25N. Computer simulations are done on this model and the simulation results are consistent with the results of theoretical analyses.

This paper describes a statistical integration algorithm for color, motion and stereo disparity, and introduces a real-time stereo system that can tell us where and what objects are moving. Regarding the integration algorithm, motion estimation and depth estimation are simultaneously performed by a clustering process based on motion, stereo disparity, color, and pixel position. As a result of the clustering, an image is decomposed into region fragments. Eath fragment is characterized by distribution parameters of spatiotemporal intensity gradients, stereo difference, color and pixel positions. Motion vectors and stereo disparities for each fragment are obtained from those distribution parameters. The real-time stereo system can view the objects with the distribution parameters over frames. The implementation and experiments show that we can utilize the proposed algorithm in real-time applications such as surveillance and human-computer interaction.

In this paper, the discrimination of ultrasonic heart (echocardiographic) images is studied by making use of some texture features, including the angular second moment, contrast, correlation and entropy which are obtained from a gray-level cooccurrence matrix. Features of these types are used as inputs to the input layer of a neural network (NN) to classify two sets of echocardiographic images-normal heart and dilated cardiomyopathy (DCM) (18 and 13 samples, respectively). The performance of the NN classifier is also compared to that of a minimum distance (MD) classifier. Implementation of our algorithm is performed on a PC-486 personal computer. Our results show that the NN produces about 94% (the confidence level setting is 0.9) and the MD produces about 84% correct classification. We notice that the NN correctly classifies all the DCM cases, namely, all the misclassified cases are of false positive. These results indicate that the method of feature-based image analysis using the NN has potential utility for computer-aided diagnosis of the DCM and other heart diseases.

A motion compensation LSI for realtime MPEG1/H.261 video encoding has been developed. This LSI employs a compact motion estimator that consists of vector search array processors. Furthermore, an efficient motion vector search strategy that enables bidirectioanl searches with a -16.0/+15.5 pels range is adopted to maintain encoded picture quality. The adopted strategy takes two steps. The first step is the full search for 2-pel precision vectors within the range of 16 pels. A 4-to-1 sub-sampling technique with a low pass filter is employed in this step. The second step is the full search for half-pel precision vectors within a 1.0 pels search range centered on the location pointed by the best 2-pel precision vectors. This strategy is compared with the exhaustive-search strategy. It is shown that the number of operations and external memory access cycles are reduced to 1/11 and 1/2, respectively, while differences of the signal to noise ratios obtained by simulation are within 0.2 dB. Those reductions contribute to lowering power dissipation. The array processors calculate the values of distortion. They accumulate the absolute differences between current and reference data with a feedback loop to keep the number of processor elements equal to the number of pels in a row of the current block. Multiple reference data buses and a delay line in the feedback loop have been introduced for efficient calculation. In addition, cascade connection of the array processors is studied to shorten calculation periods. This LSI controls input frames reordering buffers and reference frames buffers. It generates the prediction and the prediction error blocks as well as the motion vectors. AC power of current blocks and the values of distortion are obtained for the bit rate control. This LSI is fabricated using 0.8 µm 2-level metal CMOS technology and dissipates 2.0 W from 5 V supply at 36 MHz.

This paper will describe an overview on several design issues and solutions for the realization of MPEG2 encoder &decoder LSIs. ULSI technology and video-coding specific design have been able to actualize an MPEG2 encoder &decoder LSI with realtime capability, flexibility and cost effectiveness, though MPEG2 processing at MP＠ML (Main Profile and Main Level) requires an enormous computation power of 10-200 GOPS depending on the motion estimation algorithm and a search range. Video coding processors, whose performance has been enhanced at the rate of one order per 3 years, have reached the performance level required to implement MPEG2 encoding using multiple chip configuration. This has been achieved by a hybrid architecture with video-oriented RISC and hardware engine optimized for coding algorithms. Intensive circuit optimization was carried out for transform coding such as DCT and predictive coding with motion estimation. Now cost effective MPEG2 decoders have begun to penetrate the multimedia market. There are two main design issues. One is the architectural and circuit design which minimizes the silicon area and power dissipation. The other is external DRAM control which makes use of DRAM storage and band width efficiently to reduce the system cost. Also future trends in a deep submicron era will be discussed. A single chip MPEG2 MP＠ML encoder is expected to appear in the 0.25 micron era at the latest. An MPEG2 MP＠ML decoder could be compressed to an area of about 25 mm2.

This work deals with thermal-noise modeling for silicon vertical bipolar junction transistors (BJTs) and relevant integrated circuits (ICs) operating at low currents. The two-junction BJT compact model is consistently derived from the thermal-noise generalization of the Shockley semiconductor equations developed in work which treats thermal noise as the noise associated with carrier velocity fluctuations. This model describes BJT with the Itô non-linear stochastic-differential-equation (SDE) system and is suitable for large-signal large-fluctuation analysis. It is shown that thermal noise in silicon p-n-junction diode contributes to "microplasma" noise. The above model opens way for a consistent-modeling-based design/optimization of bipolar device noise performance with the help of theory of Itô's SDEs.

This paper presents an accurate and semi-physical MOSFET substrate current model suitable for analog circuit simulations. The proposed model is valid over a wide range of the electric field present in MOSFET devices and is continuous from cut off region to saturation region. The developed model was implemented into the circuit simulator, SPICE3. Benchmark of the developed model was achieved by making comparisons between the measured data and the simulated data for MOSFET devices, push-pull CMOS inverters, a regulated cascode CMOS operational amplifier. The experimental results showed that the developed model was more accurate and computationally efficient than the conventional models.

The aliasing probability was analyzed for MISRs when the error probability for each input was different. A closed form expression was derived by applying the complete weight distributions of linear codes over a Galois field and its dual codes. The aliasing probability for MISRs characterized by non-primitive polynomials was also analyzed. The inner product for binary representation of symbols was used instead of multiplication over a Galois field. The results show the perfect expression for analyzing the aliasing probability of MISRs.

An MPEG2 video decoder LSI fully compliant with MPEG2 main profile at main level is described. The video decoder LSI is a single chip solution which can implement MPEG2 video decoding with conventional DRAMs. The LSI features an architecture based on dedicated decoding hardware so as to gain the necessary computational power for real-time processing of ITU-R R.601 size video. The variable length decoder (VLD), owing to our "one symbol decoding in one cycle" policy and a special circuit for detecting unique startcodes, achieved bitstream decoding up to 18 Mbps with a normal decoding process. It also realized fast searching for the next start-code in the picture skipping and error recovery processes. The video decoder LSI also features a hierarchical and adaptive control mechanism. This control mechanism decreases the dead time of the decoding circuits and raises the efficiency of data transfer via the local DRAM port. It also contributes to the realization of error concealment and error recovery processes. This chip is capable of processing NTSC-resolution video depicted in MPEG2 MP＠ML in real-time at 27 MHz operation. The chip integrates about 1200 K transistors using 0.5 µm double metal CMOS technology. The feature of the hardware based architecture results in a low power dissipation, and the chip consumes a 1.4 W of power at 3.3 V supply voltage and is housed in a plastic QFP.

We have developed a method to automatically generate a multi-input-adder circuit for an irregular array of partial products. "FASTOOL," an FIR Filter Automatic Synthesis TOOL for an HDL design environment, is proposed for use with this method and with conventional filter coefficient design programs. Filter design from specifications to the structure of Verilog-HDL has been automated. It is possible for a system designer to quickly perform filter LSI optimization by balancing cost and performance.

This paper proposes a new method that can robustly recover 3D structure and 3D motion of 3D moving objects from a few multi-views. It recovers 3D feature points by obtaining intersections of back-projection lines which are connected from the camera's optical centers thorough projected feature points on the image planes corresponding to the different cameras. We show that our method needs only six views to suppress false 3D feature points in most cases by discussing the relation between the occurrence probability of false 3D feature points and the number of views. This discussion gives us a criterion to design the optimal multi-camera system for recovering 3D structure and 3D motion of 3D moving objects. An experimental multi-camera system is constructed to confirm the validity of our method. This system can take images from six different views at once and record motion image sequence from each view over a period of a few seconds. It is tested successfully on recovering the 3D structure of Vinus's plaster head and on recovering the 3D structure and 3D motion of a moving hand.

This paper presents some tighter bounds on universal noiseless coding, in particular, the lowerbound tighter than Davisson et al.'s for finite sequence and the upperbound for some typical universal data compression. We find that Davisson et al.'s bound satisfies some optimization in the case of using the Jeffreys prior and also that the derived upperbound in this paper is within O(1/n) from the Clarke and Barron asymptotics in the case of some restricted typical universal data compression defined in the paper.

There have been many developments on neural network research, and ability of a multi-layered network for classification of multi-spectral image data has been studied. We can classify non-Gaussian distributed data using the neural network trained by a back-propagation method (BPM) because it is independent of noise conditions. The BPM is a supervised classifier, so that we can get a high classification accuracy by using the method, so long as we can choose the good training data set. However, the multi-spectral data have many kinds of category information in a pixel because of its pixel resolution of the sensor. The data should be separated in many clusters even if they belong to a same class. Therefore, it is difficult to choose the good training data set which extract the characteristics of the class. Up to now, the researchers have chosen the training data set by random sampling from the input data. To overcome the problem, a hybrid pattern classification system using BPM and Kohonens feature mapping (KFM) has been proposed recently. The system performed choosing the training data set from the result of rough classification using KFM. However, how the remotely sensed data had been influenced by the KFM has not been demonstrated quantitatively. In this paper, we propose a new approach using the competitive weight vectors as the training data set, because we consider that a competitive unit represents a small cluster of the input patterns. The approach makes the training data set choice work easier than the usual one, because the KFM can automatically self-organize a topological relation among the target image patterns on a competitive plane. We demonstrate that the representative of the competitive units by principal component analysis (PCA). We also illustrate that the approach improves the classification accuracy by applying it on the classification of the real remotely sensed data.