Search

In this paper, the authors propose a high-speed implementation of a parameter filter, using a multidimensional space search algorithm with the introduction of virtual bounding rectangles. The authors have proposed and implemented a KUPF architecture for the problem of packet classification reflecting the complex policies of a router. Because parameter filters generally have filters for multiple parameters, the conventional methods for increasing speed are not applicable to KUPF. R-tree and R*-tree, which construct a search tree using minimum bounding rectangles and accelerate the search, are known search methods for multidimensional spaces. Meanwhile, this proposal represents filter rules in multidimensional space with the introduction of virtual bounding rectangles, while normalizing differences in the properties of each parameter and applying an R*-tree. KUPF-VR, in which this proposal is implemented, had an improved search speed compared to conventional KUPF and was clearly superior in the case of a large number of filter rules, compared to methods specialized for IP packet processing. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(13): 92–103, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20287 [ABSTRACT FROM AUTHOR]

Recently, the application of human facial shapes has been attracting interest in such fields as authentication, amusement, and medical care. However, the systems now used in practice heavily burden the person undergoing measurement, because a long time is required for measurement and the user must maintain his or her facial expression. The paper considers the measurement of the human face and proposes a fast method for head shape measurement in which the number of projected patterns is decreased by utilizing the continuity of the facial shape and the uniformity of its reflectivity. Specifically, based on the assumption of continuity of the facial shape, the pattern projection of the upper bits is omitted in the spatial coding, and the spatial code is restored by utilizing the continuity of the facial shape. By applying this method, shape measurement with 8-bit spatial resolution in the depth direction is achieved by five pattern light projection operations. Based on the fact that the reflectivity of the face is almost uniform, binarization is performed without complementary pattern projection, by using average image comparison and local threshold setting. With these measures, the head model can be acquired by the proposed method in 14 pattern imaging operations, compared with 36 pattern imaging operations required in the conventional measurement system based on spatial coding using complementary pattern light projection. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(13): 82–91, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20377 [ABSTRACT FROM AUTHOR]

A new algorithm for object classification based on an extension of Fisher's discriminant analysis is presented. Object recognition algorithms using the standard Fisher's algorithm, such as the Fisherface, train the classifier using sample-class pairs, where, for the classes, object categories determined in the application systems are used directly. In contrast, the new algorithm automatically produces subclasses, within each predetermined category, that are actually used for classification, via unsupervised learning. In order to perform this, we combine Fisher's discriminant analysis with the Akaike Information Criterion, optimizing the class configuration, that is, sample-subclass correspondences, and the feature extraction function simultaneously, thereby improving the potential of class separability. By applying this new method to face recognition, we show how it outperforms the traditional Fisher-based method. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(13): 72–81, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20378 [ABSTRACT FROM AUTHOR]

This paper discusses a modeling technique for visual kansei (visual impression), which differs depending on the individual user. The visual impression is considered at two levels, the impression at the physiological level, which does not differ much between individuals, and the impression at the psychological level, which differs according to the individual's knowledge and experience. The modeling technique for the impression at the physiological level is considered as follows. The image feature parameters simulating the feature extraction mechanism of the visual perception process are defined and are extracted from the image. A hierarchical classification of the database simulating the human intuitive image classification process, image structure estimation using MDL, and discrimination analysis are combined to produce a modeling technique for the impression at the psychological level. After combining the modeling techniques at the physiological and psychological levels, the method is applied to a similar image retrieval system. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(13): 58–71, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20400 [ABSTRACT FROM AUTHOR]

The log-polar transformation is a model of the foveal sensor, which is often seen in the visual apparatus of living organisms. The log-polar transformation is widely known in the field of image recognition, and has often been used for facial image recognition. However, this detection method is limited to 2D environments, since it is invariant only to scaling and rotation. This paper proposes a new log-polar transformation which also detects affine transformation parameters. The proposed method can detect the posture of a 2D image in a 3D environment, as well as scaling and rotation in a 2D environment. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(13): 49–57, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20564 [ABSTRACT FROM AUTHOR]

In studies of similar character recognition, the quadratic compound discriminant function has been proposed, in which a quadratic nonlinear transformation is applied to the compound discriminant function. It has been shown that the method can achieve high discrimination power. It may be true that a better decision boundary is estimated by raising the order of the nonlinear transformation, but the realization of third- or higher-order nonlinear compound discriminant functions is difficult due to the explosive increase of the amount of computation. On the other hand, the nonlinear transformation using kernel functions has recently been attracting attention. The high-order nonlinear transformation can be realized by this approach with a small amount of computation. It is expected that the above problem will be solved by performing a nonlinear transformation of the compound discriminant function by this method. This paper shows that the nonlinear transformation of the compound discriminant function can be formulated by using a kernel function, and proposes the nonlinear compound discriminant function. A recognition experiment shows that greater discrimination power than the quadratic compound discriminant function can be acquired by the proposed method with almost the same amount of computation as that using the compound discriminant function. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(11): 36– 48, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20530 [ABSTRACT FROM AUTHOR]

Three-dimensional pattern recognition requires the definition of a similarity measure between 3-dimensional patterns. We discuss how to match 3-dimensional patterns, which are represented by a set of images taken from multiple directions and approximately represented by subspaces. The proposed method is to calculate the canonical angles between two subspaces. In order to differentiate an object and its photograph, at least a canonical angle other than the smallest one is required. We discuss the usage of the canonical angle both used by itself and combination, sum and product. We also demonstrate the effectiveness of the proposed method by experiments. © 2007 Wiley Periodicals, Inc. Syst Comp Jpn, 38(9): 11–20, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20786 [ABSTRACT FROM AUTHOR]

This research paper proposes an omnidirectional pointing device that uses EMG signals. Since this device does not keep the user in front of the mouse or keyboard, there are expectations for its use as an input device for a wearable computer, and other equipment. A Recurrent Log-Linearized Gaussian Mixture Network is introduced in order to make use of the time series characteristics of EMG signals. This neural net is configured based on the hidden Markov model, which is a dynamic probability model, and infinitely many movement directions of a pointer can be expressed as combinations of probabilities of movement in preset reference directions. As a result of a verification experiment, it was confirmed that the pointer could be controlled with good precision. The effectiveness of the proposed method was confirmed by comparative experimentation with conventional methods. © 2006 Wiley Periodicals, Inc. Syst Comp Jpn, 37(4): 55–63, 2006; Published online in Wiley InterScience (www.interscience.wiley. com). DOI 10.1002/scj.20401 [ABSTRACT FROM AUTHOR]

An analytical algorithm based on neural networks is proposed for the problem of discrimination of seven subjects into the two groups of old persons and young persons on the basis of acceleration data during walking. An old person model and a young person model are constructed using neural networks, and subjects are discriminated by comparison of the degree to which the subject data match the models. To improve discrimination accuracy, a method is further proposed in which data that clearly manifest a difference in the degree of matching are added to learning data, and model reconstruction is iterated. Frequency analysis is also used to extract and quantify a feature from data on old persons and young persons and to discriminate this characteristic quantity by comparison of the degree of similarity between subjects and old persons/young persons. Validity of the discrimination results based on neural networks is examined by comparison with the results of frequency analysis. © 2006 Wiley Periodicals, Inc. Syst Comp Jpn, 37(4): 1–10, 2006; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20499 [ABSTRACT FROM AUTHOR]

This paper proposes a method for extracting text portions occurring in video images with high accuracy for reading by OCR. Past studies have produced methods of extracting text in a video image from its background by binarization based on a threshold, utilizing the fact that the intensity of the text is higher than that of the background. One method for determining the threshold is Shio's application of Otsu's method, assuming the distribution of two intensities of the background and characters in local blocks. However, methods based on various intensities of the background such as those of video images have the problem of not yielding a good threshold due to assumptions that are not necessarily valid. In addition, in reality, they cannot extract characters with accuracy sufficient for OCR readability because the intensity around the characters is not necessarily high due to the effects of shadowing, edge elimination, and signal conversion processing. Thus, this paper proposes a method of extracting only the text portions by robustly estimating the intensity distribution of the text portions, initially extracting high-reliability areas as text portions, and extending the areas based on the estimated distribution. Experimental results show that the proposed method extracts text portions with higher accuracy and better OCR readability than the conventional methods. © 2005 Wiley Periodicals, Inc. Syst Comp Jpn, 36(9): 87–96, 2005; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10148 [ABSTRACT FROM AUTHOR]

To estimate distribution from training samples, maximum likelihood estimation treats the unknown parameter of the distribution as a constant. On the other hand, Bayesian estimation treats the parameter as a random variable. In pattern recognition, Bayesian estimation has been known to improve recognition accuracy. However, it was pointed out that Bayesian estimation is not effective due to the bias of the likelihood when sample sizes of classes are not the same. In this paper, we show that recognition accuracy is improved by modifying the bias of the likelihood when sample sizes are not the same. This indicates that the cause of the ineffectiveness is the bias. We derive the formula of the bias of Geisser's predictive distribution without any approximation, and show a way of modification of the bias of the likelihood. We confirm the effectiveness of the proposed method in experiments. In addition, the derived formula gives the theoretical background of a known empirical knowledge. © 2005 Wiley Periodicals, Inc. Syst Comp Jpn, 36(5): 45–54, 2005; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.20207 [ABSTRACT FROM AUTHOR]

The research group of which the authors are members has proposed a scheme based on finger vein patterns as a scheme of biometric identification utilizing biological information. Since the finger vein images taken to obtain finger vein patterns are obtained by irradiating the fingers with infrared rays, fluctuations in brightness due to variations in the light power or the thickness of the finger occur. This paper proposes a scheme for extracting global finger vein patterns by iteratively tracking local lines from various positions to robustly extract finger vein patterns from such unclear images. The proposed scheme is robust against brightness fluctuations as compared with the conventional feature extraction schemes and the results of evaluating the proposed scheme by applying it to a personal identification system showed an effective error rate of 0.145%. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(7): 61–71, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10596 [ABSTRACT FROM AUTHOR]

We propose a new method to recognize monetary amounts printed by checkwriters. We applied a character segmentation method robust to noise and a bottom-up parsing for postprocessing. In the proposed method, the input image is segmented both horizontally and vertically in the presegmentation step. A context-free grammar describes the linguistic constraints in the monetary amounts. We devised a new bottom-up parsing technique that interprets the results of character classification. We tested the validity of the new method using 1314 images. The correct ratio was 72.7%, which is 8.4 points higher than that of the conventional method. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(7): 49–60, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10616 [ABSTRACT FROM AUTHOR]

The authors performed research on a pattern classifier using ECOC (Error Correcting Output Coding) which encodes learning patterns using error correction coding and then learns such that each discriminant function outputs the encoded values. In particular, in this research the authors focused on a method to decode the output of each discriminant function to class symbols, and then proposed a decoding method: WLS-ECOC (ECOC combination strategy based on Weighted Least Squares). WLS-ECOC estimated posterior probability for each class using the linear sum of the output of each discriminant function and then performed pattern recognition based on these posterior probabilities. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(4): 9–18, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10549 [ABSTRACT FROM AUTHOR]

Studies of immune systems include the work of Jerne on the equilibrium state of a network modeling an immune network. The authors have previously studied the information processing capabilities of immune systems and have proposed models capable of acquiring immune memory based on immune responses in an immune system. These models are based on the interactions between cells constituting an immune system, and these interactions are controlled by the state of a network formed by the cells in advance. However, in these models, cell groups are constructed by repetitions of immune responses in the immune system of a real biological organism and the fact that the immune responses are made by cell groups is not taken into account. Therefore, in this paper an immune system model involving an interactive process of network generation by cell groups generated on the basis of the clonal selection theory is proposed. The immune properties of the proposed network are shown by simulation and its effectiveness is demonstrated by applying it to a pattern recognition system. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(12): 56–63, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10243 [ABSTRACT FROM AUTHOR]

In this paper we propose an automatic recognition system for wild flowers. Two photos, one of the flower and one of the leaves taken from directly above or at a close angle of a single wild flower, were used as a single set. The objective (flower, leaf) is extracted from each image using a clustering method, then recognition is performed using a piecewise linear discriminant function after finding 10 features in the picture of a flower and 11 features in the picture of a leaf. We performed experiments on 20 sets of 34 species of wild flowers that grow around our university campus in the spring and early summer. The results of the experiment showed a recognition rate of 96.0% when all of the features (21) were used. Next we performed an experiment to select the features particularly useful for recognition. These results showed that six features for the pictures of flowers and two features for the pictures of leaves were the most effective, and with them a recognition rate of 96.8% could be reached. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(10): 90–101, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10099 [ABSTRACT FROM AUTHOR]

We propose a new registration algorithm for overlaying a virtual object generated in a graphics system onto a real-world image sequence captured by an uncalibrated camera for augmented reality. In our method, a Projective Grid Space (PGS) is defined by using two base images captured from different positions with uncalibrated cameras. In the PGS, correct perspective views of a virtual object can be registered onto an image sequence captured by an uncalibrated camera in real time. Since our method is based on the projective geometry that is defined by two base images, the registration does not require any markers with known 3D positions, but several (at least eight) image feature points are needed for estimating fundamental matrices of the augmented camera with the base images. Based on this method, we constructed an augmented reality system for an uncalibrated camera without any 3D position information in the real world, which runs nearly at video rate (11 frames per second). Experimental results demonstrate the advantages of this method. © 2003 Wiley Periodicals, Inc. Syst Comp Jpn, 34(5): 47–55, 2003; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10261 [ABSTRACT FROM AUTHOR]

In this paper, pattern recognition is reconsidered in geometrical terms, and issues in learning theory that require further research are discussed. First, it is shown that there is learning that is easy for humans and learning that is difficult for humans, which has not been duly taken into account in existing learning theories. Then the importance of pattern structure in learning is demonstrated. For the purpose of development of a new learning theory, the structure of pattern recognition is revised, and all problems of recognition are shown to have a geometrical structure, thus being treatable by means of transformation groups. In closing, issues are put forward that should be studied in the framework of new learning theory with regard to pattern structure, and coordination between theory and experiments in pattern recognition is discussed. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(13): 103–110, 2002; Published online in Wiley InterScience (www.interscience. wiley.com). DOI 10.1002/scj.1177 [ABSTRACT FROM AUTHOR]

The original “spreading associative neural network,” which imitates the spatial recognition system in the brain, not only recognizes the shape of an object irrespective of its position, but also recognizes its position irrespective of its shape in the input pattern when the object positional movement is limited to only one dimension. In the present study, in order to allow an object presentation in any 2D position on the plane, we propose a “2D spreading associative neural network” which spreads the input pattern in two dimensions. The proposed neural network not only recognizes the shape of the object in the 2D space irrespective of its position, but also recognizes its position irrespective of its shape in the input pattern. The position-invariant shape recognition and the shape-invariant positional recognition ability in 2D space are obtained across various kinds of objects—characters and human faces—at the same time. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(13): 11–23, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10171 [ABSTRACT FROM AUTHOR]

The recognition of motions is very important for understanding dynamic actions and extracting events from video sequences. In this paper, we propose invariants on object motions, which can be extracted from image sequences, and are invariant even if the motions are viewed from arbitrary viewpoints. We call them space-time invariants since they are defined under space-time projections. We show that these invariants are very useful for extracting specific motions from video sequences. The proposed space-time invariants are implemented and tested on real image sequences. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(12): 38–46, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10147 [ABSTRACT FROM AUTHOR]

Decomposing an articulated object into its constituent elements is one of the basic techniques of image recognition and its applications to structural analyses or motion measurements of human and animal bodies are being studied. In this paper, a method of decomposing an object into elements based on its motion and shape while taking account of the distribution of high-curvature points of the contour of the object is proposed for motion image sequences of objects, including objects considered to have two-dimensionally moving, branchless, invariable link-length elements and long multilink structures. Since the processing is based on the shape information obtained from all of the motion image sequences by this method, an object can be divided into regions in terms of the characteristic parts of the motion, even when the contour shape for each frame may be distorted by the effects of noise and the like, and therefore the method is very effective for preliminary processing or model building in measuring a body motion. In order to show the efficacy and an applicable example of the proposed method, the kip motion, which is a bar exercise, is measured. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(8): 98–110, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.1150 [ABSTRACT FROM AUTHOR]

A method for classifying action potentials to estimate a sensory unit activity was developed. The classification of action potentials based on their waveforms involves a pattern classification problem without supervised signal. Some methods depended on an operator for specifying parameters due to the lack of a standard criterion for classification. Therefore, we have proposed an objective classification method in this paper. The performance of our method was evaluated with a simulated signal generated by a peripheral nerve model. Our method provided more appropriate classification results against the influence of the noise process than the conventional method whose reliability was reported to be high in previous studies. The classification results of our method applied to the cuff recordings suggested that there were several different stretch receptors of different characteristics, which agrees with previous reports. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(6): 44–52, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.1131 [ABSTRACT FROM AUTHOR]

A new pattern classification method called the Kernel-based Nonlinear Subspace (KNS) method is proposed. It implements a subspace method in a high-dimensional nonlinear space by a nonlinear transformation defined by kernel functions. The Support Vector Machine, a recent popular current research topic, is a nonlinear classification method employing kernel functions and has advanced classification performance. However, as the number of patterns and the number of classes increase, the problem faced is an explosive increase in the computational complexity needed for learning. Conventional subspace methods are effective classifiers of multiple classes and are fast classification techniques. But satisfactory classification performance is not achieved if the pattern distribution is nonlinear or if the dimensionality of the feature space is small compared to the number of classes. The proposed method combines the advantages of both techniques to compensate for each other's deficiencies to realize nonlinear classification of multiple classes with advanced classification performance and low computational complexity. In this paper, we demonstrate the ability to use nonlinear transforms defined by kernel functions to formulate the nonlinear subspace method, evaluate the proposed method from the perspectives of classification performance for nonlinear distributions and multiclass distributions, the stability of the classification performance with respect to parameter variations, and the computational costs needed for learning and classification, and verify the superiority of the proposed method over conventional methods. © 2001 Scripta Technica, Syst Comp Jpn, 33(1): 38–52, 2002 [ABSTRACT FROM AUTHOR]

With stereoscopic image coding, including transfer of planar image signal along with depth (disparity) data related to object's contours, and then reproduction of stereoscopic image at the receiving end, fluctuations of disparity data cause 3D distortions in the reproduced image, which are perceived as deterioration of picture quality. In this paper, such picture quality deterioration is examined in terms of influence by fluctuation parameters, object's shape, and other factors. Experimental results indicate that the detection limit for disparity fluctuation is nearly completely dictated by frequency spectrum, and the sensitivity reaches its peak around 4.5 cycles/deg. In addition, sensitivity grows with smaller object's width within viewing angle of 3 through 12 minutes; however, sensitivity curves in logarithmic plots are very similar in shape. © 2001 Scripta Technica, Syst Comp Jpn, 32(8): 20–29, 2001 [ABSTRACT FROM AUTHOR]

In this paper, we propose an enhanced fault model to increase coverage in detecting physical faults in LSI. First, as an alternative to the single stuck-at fault model, we propose the “single gate logical fault with Hamming distance 1” (SGLFH1) model, which is capable of generating test pattern sets that detect more defects from the same number of faults. The characteristics of this fault model are discussed and the effectiveness of the test pattern sets generated by it is described. Next we show that the multiple detection test pattern set for single stuck-at faults that is conventionally used to enhance defect coverage does not detect SGLFH1 faults satisfactorily. In addition, the effectiveness of the proposed model is shown by comparing the fault detection ratio of the SGLFH1 test pattern set for physical surrogate faults [1, 2] with the corresponding fault detection ratio of the multiple detection test pattern set for single stuck-at faults. © 2001 Scripta Technica, Syst Comp Jpn, 32(6): 36–44, 2001 [ABSTRACT FROM AUTHOR]

This paper proposes a structural freehand pattern identification method for the handwritten input interface of a CAD system. In this method, a handwritten curve is described as a fuzzy symbol by recognizing it as a geometrical curve primitive, then structural matching is applied to the symbol. The system consists of two parts: the fuzzy spline curve identification (FSCI) method and the fuzzy attribute graph (FAG) method. The system uses a hierarchical approach which aims at the construction of a general-purpose kind of freehand pattern identification. This paper includes the results of experiments on the basic performance of the method. © 2001 Scripta Technica, Syst Comp Jpn, 32(6): 45–54, 2001 [ABSTRACT FROM AUTHOR]

The subspace method has usually been applied to a multidimensional space (i.e., feature space) which uses features as its basis. A subspace method can also be applied to a functional space, since the subspace can be defined by an arbitrary linear space. This paper proposes the mapping of a feature space onto the Hilbert subspace so that pattern recognition can be performed by the subspace method. The proposed method has been experimentally applied to the recognition of the Japanese hiragana syllabary and has achieved a higher recognition rate than the conventional method. This improvement is based on the fact that the dimension of the common part in all of the classes of the subspaces in the Hilbert function space becomes zero. © 2001 Scripta Technica, Syst Comp Jpn, 32(6): 55–61, 2001 [ABSTRACT FROM AUTHOR]

This paper describes a pen input system which was developed to obtain subtle writing pressure data. When the writing pressure is small, a conventional electronic pen cannot collate writing pressure data sufficiently. However, the developed pen can collate sufficient data with precision. The developed pen can manipulate any high-speed signatures by means of high-speed writing pressure processing. Further, in the pen input system, writing pressure data can be obtained by using only the pen, without any tablet. The writing pressure information obtained by the developed pen input system is analyzed and applied to writer identification. © 2001 Scripta Technica, Syst Comp Jpn, 32(6): 1–12, 2001 [ABSTRACT FROM AUTHOR]

In urinary sediment examination, untypical particles may appear, even though very infrequently. Neural network with feedback training proposed in this study offers improved classification of such untypical particles, which was difficult using conventional classification algorithms. The proposed method suggests that network is optimized using classification results obtained for typical objects. The method may be easily applied to complicated patterns with multiple parameters. Operation and efficiency of the proposed method were confirmed by computer simulation. Additional validation was obtained by applying the proposed method to classification of urinary sediment images. © 2001 Scripta Technica, Syst Comp Jpn, 32(2): 11–18, 2001 [ABSTRACT FROM AUTHOR]

In this paper, we describe an algorithm designed to recognize a large number of rectangular objects that have been placed near or in contact with one another. The conventional model-matching method, which identifies an object by evaluating the consistency between its acquired image and an object model, frequently produces false results due to patterns and characters printed on the object. We propose to solve this problem by employing a new recognition method that maximizes the global consistency between a scene and an acquired image. This method is based on the fact that the correct interpretation of an entire scene results in higher overall consistency (e.g., global consistency) between the image generated from the interpretation and the actual image acquired. This recognition algorithm first generates multiple scene hypotheses of the locations and combinations of many candidate objects from a gray-scale edge image. Then, it selects the hypothesis that has the maximum global consistency with the acquired range images. This is a further development of an object recognition method that we proposed earlier and that is based on hypothesis verification utilizing model-matching with gray-scale edge images. Our method solves the problem of false recognition patterns that result from many objects being in contact with each other. The method handles the generation and verification of hypotheses as a problem of combining candidate objects, and solves this problem utilizing a genetic algorithm. In this study this proposed method successfully recognized objects 99.8% of the time in experiments using images of actual objects. This provides evidence that the method is practical for visual recognition systems in robots. © 2001 Scripta Technica, Syst Comp Jpn, 32(2): 19–31, 2001 [ABSTRACT FROM AUTHOR]

The authors discuss a simple model for pattern discrimination incorporating temporal or spatial context by feedback of the membership values output at the high-rank winner-take-all (WTA) neurons to the lower-rank pattern selection response neurons. They propose a teacherless training method based on maximum likelihood. First, for the temporal context, they study examples with the feedback of discrimination effect of a first moment to the next moment, and demonstrate that clustering is performed by proximity of presentation moments rather than by the pattern similarity. Next, for the spatial context, they show that a similar pattern recognition method can be applied to spatial smoothing of image patterns. © 2000 Scripta Technica, Syst Comp Jpn, 32(1): 45–52, 2001 [ABSTRACT FROM AUTHOR]

When only similarities or dissimilarities of two objects are given, visualization by arranging objects into two-dimensional Euclidean space is effective in representing the relationship between more than two objects intuitively. In this paper, an algorithm for arranging objects into multidimensional Euclidean space with reflection of the similarity or dissimilarity of two objects is proposed. As an application of the proposed method, classification of meanings of Japanese polysemous verbs is adopted. The effectiveness of the proposed algorithm is shown by comparison with the conventional methods. © 2000 Scripta Technica, Syst Comp Jpn, 31(13): 11–18, 2000 [ABSTRACT FROM AUTHOR]

The radial basis function network is a suitable neural network for the function approximation problem and for pattern recognition. The radial basis function network has the ability to perform local learning at each neuron, making it superior to multilayer neural networks. But the number of neurons required in order for the radial basis function network to approximate an unknown nonlinear function is not clear in advance, a fact that leads to delay and overfitting of learning. We initially proposed a competitive radial basis function network to solve these problems. The proposed competitive radial basis function network can learn effectively by means of a synapse plasticity equation that takes account of competition among the synapse weights. The competitive radial basis function network has the ability to remove redundant radial basis functions, but it cannot add necessary radial basis functions. In this paper we therefore propose an effective method for the addition of necessary neurons based on an investigation of the synapse plasticity equation. In this method, the newly added radial basis function takes over some properties of the radial basis functions that have been acquired when the learning converges. We then propose a reproductive and competitive radial basis function network, which is a neural network combining this method and the competitive radial basis function network. Simulations show that the reproductive and competitive radial basis function network can decrease the sum of the squared errors effectively in application to the function approximation problem. Our model proves able to learn faster than the usual radial basis function network. © 2000 Scripta Technica, Syst Comp Jpn, 31(13): 65–75, 2000 [ABSTRACT FROM AUTHOR]

A new algorithm based on the subspace method for recognizing poor-quality character images is proposed. In many cases, damage caused by photocopying or facsimile occurs accidentally. It is difficult to solve this kind of damage by statistical methods. We investigate how distribution of data changes because of blurring. Then the eigenvectors and eigenvalues that are used in the subspace method and multiple similarity method are modified according to the change. The effectiveness of the proposed algorithm is shown by the experimental results. © 2000 Scripta Technica, Syst Comp Jpn, 31(12): 1–10, 2000 [ABSTRACT FROM AUTHOR]

We investigate the relation between object recognition and formation of hand shape in human grasping movements through psychophysical experiments. There were few differences between the experimental data when the subjects formed the hand shape without touching objects and when the subjects actually grasped the objects. Both data included information related to the size and shape of the objects. The experimental results suggest that humans are able to compute the hand shape needed to grasp the objects from their visual properties alone before actual grasping. Moreover, developing a five-layered neural network model, we confirmed that the hand shape can be computed from only visual information in the feed-forward manner and that the property information of the objects is extracted when visual-motor transformation is performed. © 2000 Scripta Technica, Syst Comp Jpn, 31(12): 11–22, 2000 [ABSTRACT FROM AUTHOR]

This paper proposes a method for cursive handwritten word recognition. In the traditional research, cursive handwritten word recognition used a single method for character recognition. Our research proposes a method that integrates multiple classifiers in order to improve the word recognition rate by combining their results. Our experiment demonstrates that two classifiers outperform the word recognition rate of any single character classifier. © 2000 Scripta Technica, Syst Comp Jpn, 31(12): 23–32, 2000 [ABSTRACT FROM AUTHOR]

Many studies about restoring blurred images have been performed. But most of them dealt with restoration of one blurred image. However, when the object is observed several times, the restored image using the several observed images is better than the restored images using each observed image. Therefore, this paper proposes a method of restoring the several observed images of the same object. The following three processes are necessary in order to obtain a good restored image: (1) estimating the parameters of both blur process and original image model, (2) restoring via the estimated parameters, and (3) adjusting the position between the observed images. This paper deals with out-of-focus blurred images. The parameters of the blurred images are estimated by means of the characteristic of the discrete cosine transform (DCT) spectrum, and the parameters of the original image model are estimated by fitting the autocovariance function of the original image model to the sample autocovariance function of the blurred image. The iterative restoring method is derived based on the Bayes rule. In our computer simulation of both blurred images, which are generated by known blur process, and the real out-of-focus blurred images, the restored images using the several blurred images are satisfactory because the restoration using several blurred images compensates for the lost information of the images. These results are explained by the DCT spectrum. © 2000 Scripta Technica, Syst Comp Jpn, 31(9): 18–27, 2000 [ABSTRACT FROM AUTHOR]

Focusing on classification problems, this paper presents a new method for linearly combining multiple neural network classifiers based on statistical pattern recognition theory. In our approach, several neural networks are first selected, each of which works best for each class in terms of minimizing classification errors. Then, they are linearly combined and form an ideal classifier able to take advantage of the strengths of the individual classifiers, to avoid their weaknesses, and to improve all of the individual classifiers. In this approach, the minimum classification error (MCE) criterion is utilized to estimate the optimal linear weights. In this formulation, because the classification decision rule is incorporated into the cost function, better combination weights suitable for the classification objective can be obtained. Experimental results using artificial and real data sets show that the proposed method can construct a better combined classifier which outperforms the best single classifier in terms of the overall classification errors for test data. © 2000 Scripta Technica, Syst Comp Jpn, 31(9): 39–48, 2000 [ABSTRACT FROM AUTHOR]

In this paper, we propose a character processing system (undefined character processing system) that handles characters for which fonts or character code systems have not been provided on the terminals (undefined characters). The undefined characters include characters historically not installed in the input/output environment, such as ancient characters or characters used by minority groups, and characters not used in a particular language system environment. The proposed system uses morphology to analyze the structures of the target undefined characters, generates codes which describe the characters' shapes based on the analysis results, and reconstructs the characters from these codes. A prototype of this system was constructed. The results of the structural analysis of the undefined characters and their reconstructions verified the effectiveness of the proposed system. © 2000 Scripta Technica, Syst Comp Jpn, 31(9): 60–71, 2000 [ABSTRACT FROM AUTHOR]

There exist methods of searching for an object whose position and size are not known in an image. To obtain an accurate result by such methods, the focused region of an input image has to be compared many times with its reference image by varying its position and size. This makes fast searching difficult. This paper describes a method of fast active search for an object and its position in an image by using the color histogram, which is stable to variations of shape of the object. When the input image of an object is compared with its reference image, the upper limit of the similarity in the image region is calculated from the input image in a certain position using the algebraic characteristics of the color histogram. If the upper limit is smaller than the search value, no search in this region is necessary, so that the number of matching operations can be greatly reduced. This method involves no approximations and has a computation speed about 10 to 1000 times faster than conventional exhaustive search methods. Applications of the proposed method, such as tracking, retrieval, and counting, are also described. © 2000 Scripta Technica, Syst Comp Jpn, 31(9): 81–88, 2000 [ABSTRACT FROM AUTHOR]

We have developed a high-speed pattern recognition system using evolvable hardware (EHW). EHW is hardware that can change its own structure by genetic learning for maximum adaptation to the environment. The purpose of the system is to show that recognition devices based on EHW are possible and that they have the same robustness to noise as devices based on an artificial neural network (ANN). The advantage of EHW compared with ANN is the high processing speed and the readability of the learned result. In this paper, we describe the learning algorithm, the architecture, and the experiment involving a pattern recognition system that uses EHW. We also compare the processing speed of the pattern recognition system with two types of ANN dedicated hardware and discuss the performance of the system. © 2000 Scripta Technica, Syst Comp Jpn, 31(4): 1–11, 2000 [ABSTRACT FROM AUTHOR]

This paper describes a new duration-based state transition model that incorporates a discriminative weighting parameter. The serious problem of acoustical/temporal variation in different speakers' utterances has long made it difficult to design a robust speech recognition model. To cope with this problem, we integrate state-dependent weighting parameters into the structure of the duration-based state transition model and control the score for the accompanying temporal segment. These state-dependent weightings absorb local variations in an utterance and provide discriminative ability that can directly influence the total score for the speech pattern. Moreover, a duration-based matching procedure precludes stiff dynamic warping that involves excessive time stretching or compression. We train all of the model parameters including proposed parameters, by using a minimum classification error criterion. Minimum classification error training is implemented by a generalized probabilistic descent method that can directly minimize the number of recognition errors between competing models. We also discuss the relationship of this model to a conventional hidden Markov model, and applicability of the minimum-classification-error/generalized-probabilistic-descent training to the hidden Markov model. The experimental results of spotting-based speaker-independent word recognition show that both state-dependent weightings and duration control helped to improve recognition performance. Furthermore, for low-performance speakers, the recognition rate was raised from 65.5% to 92.0% after speaker adaptation using the training, proving that the training was effective. © 2000 Scripta Technica, Syst Comp Jpn, 31(4): 98–108, 2000 [ABSTRACT FROM AUTHOR]

Many relics from ruins are small fragments and work is required to restore their original shapes in order to research, analyze, or display them. However, such work is very difficult, and there are problems associated with handling the relics without causing damage and not being able to disassemble the relics once restored, due to the use of fixing agents and the like. The purpose of this paper is to construct a virtual reality environment inside a computer, reconstruct the three-dimensional shape of a relic inside the virtual reality environment, and construct an environment for restoring the relic without using the original relic. The three-dimensional shape of the fragment is measured using MRI. The image data thus obtained are converted into polygon data by image processing to be reconstructed in the virtual reality environment. The fragments reconstructed inside the virtual reality environment are put into their original shape by assembling them using a data glove. © 2000 Scripta Technica, Syst Comp Jpn, 31(4): 52–62, 2000 [ABSTRACT FROM AUTHOR]

In order to enhance the accuracy of mobile robot navigation, we propose a new method of pattern matching to estimate the robot configuration. It matches the observed data acquired by an environmental sensor to a given map of the robot environment. In this paper, we assume that the environment is expressed by a set of line segments in a two-dimensional space. We also assume that the environmental sensor consists of a set of range sensors oriented radially from a point, and that each range sensor has a normal error distribution. Our algorithm minimizes the expectation of the error of matching. The three-dimensional configuration space is divided into perceptual equivalence classes, then the minima in each class are sought in the x-y plane and in the θ direction. Experimental results are shown. © 1999 Scripta Technica, Syst Comp Jpn, 31(1): 30–38, 2000 [ABSTRACT FROM AUTHOR]

We are trying to realize gesture recognition from motion images that capture human gestures and hand motions without the need for tactile sensors or markers, such as data gloves. We have achieved real-time spotting-based recognition with the continuous DP system without any gesture performers having to be aware of the commencement and ending of a motion. Continuous DP is beneficial due to its ability to create a model from a single-gesture motion image sequence. Therefore, even if gesture motions depend on performers, gestures can be recognized after a single instruction. However, to instruct a model, our conventional real-time recognition systems require setting of the threshold for recognition as well as segmentation of the model off-line. Therefore, we propose in this paper an automatic model segmentation method and a recognition method to allow constructing performer-adaptive on-line teaching systems. We will also demonstrate the usefulness of this method through evaluative experiments. © 1999 Scripta Technica, Syst Comp Jpn, 31(1): 39–47, 2000 [ABSTRACT FROM AUTHOR]

One of the typical filtering approaches, in which no constraint is imposed on the document structure, is the method based on the vector space model. In this method, the words composing the document are assigned to the elements of the vector, and the document is selected statistically by the similarity in the constructed feature vector space. Hitherto, the documents have been ranked in the order of high similarity, and the one-dimensional filtering is considered, based on the specified number of documents in the upper rankings or a preset threshold. A problem in this filtering is that only the necessary information is selected following the order, and noise elimination to reject the unnecessary information is not directly considered. From such a viewpoint, this study takes the approach of considering the user's interest, both positive and negative. The filtering is considered as the corresponding two-category problem, and a filtering method is proposed based on the concept of the pattern classification. This method uses the subspace method, which is known as a method for pattern recognition. The proposed method consists of highly precise filtering introducing the co-occurrence relation among words, and realizes the representation and updating of the interest items by a single mechanism. The effectiveness of the proposed method is demonstrated by experiments using newspaper articles. © 1999 Scripta Technica, Syst Comp Jpn, 31(1): 48–58, 2000 [ABSTRACT FROM AUTHOR]

Nearest neighbor pattern recognition is represented in terms of robust estimation, and a classification method using multimodal data fusion based on the Bayes rule is proposed. The proposed method is proved to be a kind of fuzzy voting. Unsupervised learning of classes' representative points using the EM algorithm is introduced. The basic properties of the proposed multimodal classifier are examined using simple data, and a qualitative explanation of the McGurk effect is offered. Experimental results on segmentation of multiple images are presented as an example of application. © 1999 Scripta Technica, Syst Comp Jpn, 30(8): 51–60, 1999 [ABSTRACT FROM AUTHOR]

There have been many attempts to identify an unknown input pattern from patterns having many categories represented by multiple-dimension feature vectors, but they require a very large computation time when a conventional method is applied. This article proposes a high-speed automatic method of determining a small number of likely categories without calculating distances to standard patterns. In the proposed method, the region of existence of samples on each feature axis is defined by using the learning-sample distribution, and the region is divided into l cells. Then a dictionary is created so that a category is given when a class is specified. The class is determined from the feature elements of an unknown input pattern. This procedure is applied to all of the feature elements so that the number of cumulative features is obtained. By taking the cumulative features from the maximum to the (maximum – k) (an integer) and applying postprocessing, a likely category is determined. The effectiveness of the proposed method has been confirmed by using 1153 on-line handwritten characters (Japanese phonetic letters and Chinese characters). © 1999 Scripta Technica, Syst Comp Jpn, 30(9): 34–43, 1999 [ABSTRACT FROM AUTHOR]

A neural network model that recognizes sequential patterns without expanding them into spatial patterns is presented. This model forms trajectory attractors in the state space of a fully recurrent network by a simple learning algorithm using nonmonotonic dynamics. When a sequential pattern is input after learning, the network state is attracted to the corresponding learned trajectory and the incomplete part of the input pattern is restored in the input part of the model; at the same time, the output part indicates which sequential pattern is being input. In addition, this model can recognize learned patterns correctly even if they are temporally extended or contracted. © 1999 Scripta Technica, Syst Comp Jpn, 30(4): 11–19, 1999 [ABSTRACT FROM AUTHOR]

A neocognitron is a hierarchical neural model which can recognize shifted patterns in positions and/or deformed patterns after unsupervised learning, but the model cannot recognize rotated patterns. We propose a rotation-invariant neocognitron which can also recognize rotated patterns by extending the function of the neocognitron. We also propose a new learning method, by which thresholds of the rotation-invariant neocognitron are controlled in a training phase so that the model shows the expected performance for the recognition of rotated patterns. We show that the model can recognize rotated patterns as well as shifted and/or deformed patterns in computer simulations. The model also acquires an ability of recognition of locally rotated patterns, which cannot be correctly recognized by the standard model, because the rotation-invariant neocognitron recognizes the entirety of an input pattern through multilayered processing of local features including rotational information. © 1999 Scripta Technica, Syst Comp Jpn, 30(4): 31–40, 1999 [ABSTRACT FROM AUTHOR]