Recent progress in large-vocabulary continuous speech recognition (LVCSR) has raised the possibility of applying information retrieval techniques to the resulting text. This paper presents a novel unsupervised text segmentation method. Assuming a generative model of a text stream as a left-to-right hidden Markov model (HMM), text segmentation can be formulated as model parameter estimation and model selection using the text stream. The formulation is derived based on the variational Bayes framework, which is expected to work well with highly sparse data such as text. The effectiveness of the proposed method is demonstrated through a series of experiments, where broadcast news programs are automatically transcribed and segmented into separate news stories. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 1–11, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20421

This paper proposes a method of obstacle detection using vehicle-mounted stereo cameras. There are various objects in the road environment; these can be divided into objects, such as pedestrians, which have the possibility of colliding with the vehicle, and objects, such as road patterns, which do not have such a possibility. In this paper, an obstacle is defined as an object which is located in the space through which the vehicle will travel. In order to decide whether an object is an obstacle from the pattern in the image, it is necessary to estimate the position of the pattern in relation to the road surface. For this purpose, the planar region is determined by dynamically estimating the homography (projection transformation) matrix between the images corresponding to the plane. The pose of the camera with respect to the road is determined. Then the spatial position obtained by stereo measurement is converted to road coordinates, based on the orientation of the road surface. By applying segmentation based on the converted result of stereo measurement, obstacles are detected. On ordinary roads it is required to detect various obstacles, ranging from relatively small obstacles such as pedestrians and bicycles to large obstacles such as cars. The proposed method can detect obstacles without explicit modeling. The effectiveness of the proposed method is demonstrated by experiments using various actual images. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 12–22, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20413

We propose a statistical regression testing method for evaluating the reliability of software as part of the software maintenance process. Maintenance procedures take up more than half the time of the software development process; in addition, software reliability is an important factor in determining the dependability of a product. Regression tests are performed in order to conserve or improve software reliability as part of the software maintenance process. However, existing systematic testing methods based on regression tests are not necessarily appropriate for evaluating software reliability. The statistical regression testing method is a means for compensating for the flaws of such existing methods. In this method a model of how the user makes use of the software is defined by means of a Markov chain; this is known as the usage model, and then test cases are generated at random according to a probability distribution based on this usage model. In this paper we perform experiments applying the proposed method to a small-scale client server program and demonstrate that the proposed method can be implemented. In addition, we clarify the effects and issues that may be anticipated when applying the method and establish how it may be used in practice. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 23–34, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20414

This paper proposes a new voting method using light that makes possible the realization of a high-performance computing system equipped with both flexibility and dependability. Soft errors inside computers due to terrestrial radiation have become a problem in recent years. We have proposed a highly dependable triple voting-type computing system that can be used even in the strong radiation environment of space. While this system uses reconfigurable SRAM-based FPGA to achieve high flexibility, the system operation speed was limited by synchronization problems in voting circuits located on parallel buses, and by the bottleneck of a limited upper speed for the radiation-tolerant device used in the voting logic circuit. As a result, we study the use of a serial bus, and propose a voting circuit using light that is suitable for a high-speed serial bus. This optical voting circuit is highly compatible with the high-speed serial interfaces provided on recent FPGAs, and is easily configurable using existing optical communications devices. In operations tests using an optical communications module, we obtained voting performance at a 1 Gbit/s communication speed, and estimated that the upper speed limit for the test system is 2.14 Gbit/s. Use of a higher speed optical communication module and circuit optimization should make it possible to achieve even higher speeds. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 35–46, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20422

In this paper we extend the application of neural network-based photometric stereo founded on the principle of empirical photometric stereo to color images proposing a method for computing both the normal vectors of a target object and its color reflectance coefficients. This method is able to render objects that have non-Lambert reflectance properties without using any parametric reflectance function as a reflectance model. In addition, we propose a novel neural network-based rendering method that allows the generation of realistic virtual images of an object with arbitrary light source direction and from arbitrary viewpoints based on the physical reflectance properties of the actual object and perform a comparative evaluation with approximations by existing models, the Phong model, and the Torrance–Sparrow model. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 47–60, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20423

With recent increases in criminal activity, research and development efforts are required for detection of intruders. This paper proposes an element technique for developing a cheap intruder detection system with high performance. The following three functions are developed in this study. (1) The system can deal with smooth changes of illumination, including movement of the light source. (2) For a slowly moving intruder, the region of presence is clearly detected. (3) For an intruder pausing his movement, the region of presence is clearly detected. Functions (1) and (2) are apparently contradictory. The basic idea of the method is to detect the region of presence of the intruder based on the difference image between the background image without the intruder and the present input image. The crucial point then is how to update accurately and simply the present background image, and there has been no method that realizes the above three functions at the same time. In contrast, we realized these functions by an approach in which weighted averaging is not applied to the detected intruder region and its boundary in the background image. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 61–76, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20424

This paper proposes a system in which a human image with movement is captured by a multiviewpoint camera and is converted in real time to three-dimensional data, which are sent by network transfer so that the image can be displayed at a remote site in three dimensions, using an omnidirectional display device. In the proposed system, the target human image is reconstructed in voxelwise, using parallel volume intersection for multiple images. The reconstructed voxels represent only the presence or absence of the object. Then coloring is required, based on texture information that is obtained by projecting the voxel on each image plane. The judgment as to whether the voxel is visible is important in the coloring process. In this study we attempt to achieve real-time processing from imaging to display. A parallel coloring procedure is developed by using a human image buffer in each camera, so that the visibility of the voxels is judged in real time. In order to investigate the effectiveness of the system, an experimental system for multiviewpoint image processing was constructed using five cameras and six PCs. A human image with movement was captured and a performance of 2.63 FPS was achieved for reconstruction with a spatial resolution of 96×96×96 (a voxel size of 10 mm). The reconstructed data were directly transferred to the omnidirectional display device without being stored in memory, and were displayed, demonstrating the effectiveness of the proposed method. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 77–86, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20416

In the field of life science, a tremendous amount of quantitative data related to biological systems continues to accumulate. There are still many aspects of biological phenomena which are not fully analyzed, due to interaction between various phenomena and mechanisms. The simulation technique provides useful approaches to the analysis of such phenomena, in which the model describing the phenomenon and the model parameters are adjusted. Usually, however, simulation models for biological functions are very complex, and a tremendous amount of time is required in comprehensively adjusting the model and its parameters and then evaluating the simulation results. This study aims at construction of a model for the left ventricular motion, and considers the left ventricular shape model, the cell orientation model, and the coronary artery model as its components. An interface is constructed by which the above models and the model parameters can be adjusted. Using this interface, it is possible to evaluate the simulation model efficiently. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 87–98, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20425

This paper describes a learning method of color transformation considering local illumination changes in a scene for interactive object recognition. We deal with refrigerator scenes with external and internal light. The external illumination changes are expressed in terms of locations in the refrigerator and the parameters are learned. Because the intensity of the internal illumination depends not only on the location but also on the transmissivity of target objects and existence of shading objects, we try recognizing objects by changing the intensity and determine one at which the object is successfully recognized. Experiments of object detection show the effectiveness of our method. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 99–110, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20426

This paper considers the rectangular packing problem and investigates the effectiveness of the neighborhood cultivation genetic algorithm (NCGA). NCGA, which we proposed, is a new algorithm in which the unique mechanism of neighborhood crossover is combined with effective mechanisms for search in multi-objective GA proposed in the past. The effectiveness of the proposed method in typical test problems has already been investigated with satisfactory results in past studies. The rectangular packing problem, on the other hand, is applied to floor planning, such as chip area minimization in large-scale integrated circuits. It is a kind of discrete combinatorial problem in which it is known that the search is difficult and a very long time is required until the solution is obtained. This paper formulates the rectangular packing problem as a multi-objective problem with the vertical and horizontal lengths of the placement configuration as the objectives. The sequence-pair is used as the block placement representation, and PPEX is used as the crossover procedure. Using these processes, the effectiveness of NCGA is investigated. For comparison, three other methods—NSGA-II, SPEA2, and non-NCGA (NCGA without neighborhood crossover)—are investigated. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 111–120, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20427

The purpose of this study is to establish a method for estimating the sea wave period by satellite radar altimetry. The backscatter coefficient, which is closely related to the period, is evaluated by computer simulation. Specifically, the following method is presented. The sea wave is modeled on the basis of the Pierson–Moskowitz spectrum. The average received power is determined by radar altimetry simulation of our proposal. It is shown that the backscatter coefficient of the sea surface can be calculated on a relative basis. Using the data for the relative received power calculated from various sea wave parameters, the relation between the significant wave height and the backscatter coefficient, and also the relation between the significant wavelength and the backscatter coefficient, are derived. Thus, the possibility of observing the sea wave period by using satellite radar altimetry is demonstrated. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 121–128, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20428

As a time series prediction model for sequential learning considering memory size limitations, this paper proposes the Adaptive and Sequential Learning Network (ASLN) model. The proposed model sequentially memorizes time sequence information given as the input and then performs prediction based on the memory. While effective use of the memory capacity is attempted for changes in the ambient environment, the model can follow up by varying its own memory. The model memorizes the elements contained in the input time series and the information on their transitions. It identifies the elements with a higher frequency of inputs among the time series and memorizes them as priority items. Information on the transitions of the input elements is represented as a state vector and is memorized by a layered neural network. The state vector maintains information on the past input sequence so that expression of the context is possible. A numerical experiment shows that the proposed model can predict a time series while tracking environmental changes. An experiment on learning of a number sequence was performed, using handwritten number patterns containing fluctuations. The prediction capability was verified with an increasing number of patterns. Guidelines are also provided for setting of parameters, which is important when the model memorizes the transition information of the time series. © 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(12): 129–139, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20406