3D structure of proteins deeply relates to their functionality. It is well-known that functions of proteins strongly appear in the bumpy parts of the molecular surfaces, and therefore geometric analysis of protein molecular surfaces is important.
We propose a technique to extract geometric features of protein molecular surfaces, and a visual interface to effectively visualize the extracted results. Assuming that molecule surfaces are approximated as triangular meshes, the technique extracts groups of triangles forming partial reliefs, calculates their feature values, and finally clusters them according to the feature values. In addition, our technique extracts larger similarly shaped parts consisting of two or more reliefs, by simplifying the triangular meshes and applying the graph route problem. The technique provides a visual interface for visualizing the above results applying a hierarchical data visualization technique “HeiankyoView”.

抄録全体を表示

This paper describes an attempt to build a 3D face model for an anthropomorphic interface that transmits an arbitrary combination of identity, various facial expressions, and gaze. It was done by combining a 3D morphable face model by which variations of 3D shape are represented in a small number of parameters and the Galatea face model by which facial expressions are generated. We also developed an eyeball model for visualizing eye movements to turn the gaze.

抄録全体を表示

This paper proposes a novel method for real-time rendering of polygon mesh surfaces with reflection or refraction. The basic process is similar to dynamic environment mapping or cube mapping. Our proposed method is superior to those in that the accurate ray direction is reflected in the resulted image at every vertex on the mesh. Existing real-time techniques suffer from the differences between the viewpoint for the environment map and each reflection point. The proposed method minimizes this by finding an optimal viewpoint for the reflective or refractive mesh. With a sufficient number of vertices and map image resolutions, the users can render reflected images as accurate as ray tracing for all practical purposes, except for reflected objects around ray converging points of reflection on concave surfaces or refraction through convex lenses. The method can be applied to areas which require accuracy such as industrial design. Experiments with a CAD model of a car rear-view mirror and spectacle lenses exhibited results of sufficient quality for design verification.

抄録全体を表示

In this paper, a classifier called Generalized Learning Local Averaging Classifier (GLLAC) is proposed for image classification. GLLAC is regarded as a combination of Local Averaging Classifier (LAC) and Generalized Learning Vector Quantization (GLVQ) for achieving low error rates with small amount of reference vectors. In GLLAC, all k-near reference vectors of the nearest mean vector belonging to the same class to an input vector are moved toward an input vector, whereas those of the nearest mean vector from a different class are moved away from an input vector. The performance of GLLAC is verified with experiments on handwritten digit and color image classification. Experimental results show that GLLAC can achieve lower error rates than conventional classifiers such as GLVQ or Support Vector Machine (SVM).

抄録全体を表示

The improved super-resolution method that achieves dense motion estimation (DME) using overlapped block matching (OBM) remarkably improves the quality of reconstructed images for a given video sequence. However, this method has a drawback to increase the computational cost almost linearly to the number of overlapped blocks because DME using OBM allocates multiple motion vectors to a local region in the image restoration process. To solve this problem, in this paper we propose a method to reduce computational cost of the improved super-resolution method by considering the statistics of motion vectors obtained by DME using OBM. This method can reduce the entire computational cost up to 29.9 ∼ 49.1% depending on a given video sequence while completely maintaining the original performance of the improved super-resolution method using OBM. Also, we try to further reduce computational cost by relaxing the complete original performance preservation requirement. With this additional attempt, we can further reduce computational cost up to 16.9 ∼ 20.8% without serious deterioration of the quality of reconstructed images.

抄録全体を表示

This study presents a new algorithm for automatic image-map alignment problem using a new similarity measure named Edge-Based Code Mutual Information (EBCMI) and 3-D Hilbert scan. In general, each image-map pair can be viewed as two special multimodal images, however, are very different in their representations such as the intensity. Therefore, the normal Mutual Information (MI) using the intensity in traditional alignment method may result in misalignment. To solve the problem, codes based on the edges of the image-map pairs are constructed and Mutual Information of the codes is computed as the similarity measure for the alignment in our method. Since Edge-Based Code (EBC) is robust to the differences between the image-map pairs in their representations, EBCMI also can overcome the differences. On the other hand, the 3-D search space in alignment can be converted to a 1-D search space sequence by 3-D Hilbert Scan and a new search strategy is proposed on the 1-D search space sequence. The experimental results show that the proposed EBCMI performed better than the normal MI and some other similarity measures and the proposed search strategy gives flexibility between efficiency and accuracy for automatic image-map alignment task.

抄録全体を表示

A multipose human face recognition approach is presented. The proposed scheme is based on frequency analysis (i.e. DCT or wavelet transforms) to obtain facial features which represent global information of face image and modified LDA (M-LDA) to classify the facial features to the person's class. The facial features are built by selecting a small number of frequency domain coefficients that have large magnitude values. Next, from the facial features, the mean of each face class and the global covariance are determined. Finally, by assuming that each class has multivariate normal distribution and all classes have the same covariance matrix, M-LDA is used to classify the facial features to the person's class. The aims of proposed system are to reduce the high memory space requirement and to overcome retraining problem of classical LDA and PCA. The system is tested using several face databases and the experimental results are compared to well-known classical PCA, LDA, and other established LDA (i.e. DLDA, RLDA, and SLDA).

抄録全体を表示

This paper discusses the use of structure information of Uyghur characters as a feature in online handwriting recognition on portable digital devices. Based on the position of secondary stroke, Uyghur characters can be separating into 4 groups. In this case, the unknown input character compares to other characters in its corresponding group only. This will be shortening the comparison time. The experiment result of freely-written 10 dataset showed that, the comparison time is reduced by 64.32% while the recognition rate improved 5.87%. The Approximate Stroke Sequence String Matching method is applied to Uyghur handwriting character recognition and an average recognition rate of 93.95% is obtained. It is improved to 96.6% while using the structure information of handwritten characters.
Based on these results we discuss that, the recognition rate will be improved by the using of some other features like character frequency, the number of secondary strokes, and other.

抄録全体を表示

We propose an object detection method that aims to protect privacy while providing video surveillance even when the illumination changes. Recently, many surveillance cameras have been installed in public spaces for monitoring activities. Unfortunately, there are too few authorized people to allow all of the resulting video streams to be perused continuously. Therefore, it is necessary to allow concerned citizens to watch the video streams. To permit this, we must address the privacy issue. In the proposed method, we first extract the object regions (which include private information) using the background estimated by taking account of illumination variations. Next, we filter the extracted regions to protect privacy. Experiments reveal that our method can successfully catch objects in surveillance videos, even when the objects stop for a long time under varying illumination conditions, while concealing the private information.

抄録全体を表示

This paper proposes a novel data hiding method utilizing Mquant as the data carrier in MPEG domain. To the best of our knowledge, Mquant is never used as the data carrier for information hiding. Mquant is the principle part of rate controller in MPEG, and it is one level higher than the existing data carriers such as quantized DCT coefficients and motion vectors in the MPEG coding hierarchy. In our method, matrix encoding is utilized as the data representation scheme for reducing the number of modification whenever possible. A modification scheme is proposed to sub-optimally preserve the original distribution of Mquant during data embedding. Our data hiding method is applicable not only to MPEG1/2/4 encoded video but also to the encoding process of MPEG video from a sequence of raw pictures. Carrier capacity, histogram distance, image quality, and filesize change are considered to verify the basic performance of the proposed method using various videos encoded by MPEG1. Comparisons among the proposed and existing data carriers are carried out using the same evaluation criterion. The influence of video bitrate on the performance of our method is also investigated.

抄録全体を表示

An efficient partial image storage method is required for an image revision system to store a huge amount of retouched image data. This paper proposes and evaluates a framework for an image revision management system, which uses the JPM file format to store differential elements between two versions of the image. Two data storing approaches are considered, one stores replaced elements after the edition, and the other stores replaced elements or differential elements adaptively. These approaches are evaluated with four different kinds of edit operations: blurring, edge emphasis, character writing, and adding noise. Since the combination of JPM's image object organization and JPEG 2000's efficient region compression is effective for image revision management, our proposed system can save storage capacity and network bandwidth in the image revision management system.

抄録全体を表示

We propose an image composition method which seamlessly matches the color of a source image region to that of a target image region that is partially occluded by foreground objects. Previous methods assume that a target image region has small color variation, and therefore it is difficult to paste source image regions so that they overlap foreground objects in a target image, as this induces color bleeding from the foreground objects.
To overcome this difficulty, we propose to perform color matching only from the background region by excluding the foreground objects. We show how we compose objects from a source image both behind and in front of objects in a target image, and we demonstrate that visually pleasing seamless composition can be achieved.

抄録全体を表示

In recent years, we can get high accuracy cross-sectional images with advanced medical devices such as CT, MRI, and PET, and the images are often vital to medical procedure. Images taken at different times are deformed by visceral movement. The deformations are non-rigid deformation. For medical diagnosis using the images, it is desired to develop a non-rigid image registration.
In this research, we register two images that are acquired at different time. We have developed a method for non-rigid image registration, where we use a free-form deformation for image alignment, sum of squared difference as our similarity measurement, and a steepest descent method for optimization. The method achieves improved processing speeds with user interaction for specifying deformation areas.

抄録全体を表示

We propose a system for automatic acquisition of walking persons' faces using a color video camera in conjunction with a planar laser scanner. The system reconstructs the person's body surface, estimates the location of the neck reference point in 3D space, and determines the location of this reference point on the image plane. Face images are then cut out with respect to this reference point. A method for estimating the motion velocity vector of the walking person, which is necessary for body surface reconstruction, is also proposed. A prototype system was built, and its evaluation showed promising results.

抄録全体を表示

Motion capture techniques have come to be used for digital archiving of intangible cultural properties especially traditional dances, and the data obtained have been used for movement analysis of these dances. The objective of this research is to develop and evaluate a method of segmenting the whole body motion data during dance into a number of unit movements. Segmentation of the body movement is an essential processing task for doing various quantitative and/or qualitative researches on dance body motion. The method uses information about velocity of body parts, a change of a movement direction, and a change of size of the space produced by the whole body. The method proposed has been evaluated by using ballet motion data and contemporary street dance data, and the experiments showed promising results.

抄録全体を表示

A unique feature of Jiuta-Mai in Japanese classical dance is a set of spiral movements propagated gradually from one part to another adjacent part of the body. In order to quantitatively realize the existence of spiral movement, analysis of a 3 dimensional (3D) time series of a dance motion performed by two master Japanese classical dancers was carried out. It revealed that the transformation of original 3D coordinates into a local coordinate system consisting of principal components, was effective. The first principal component represented the overall direction of body motion, and the spiral motions were clearly visualized through cyclic curves appearing in the second and third components, perpendicular to the first component.

抄録全体を表示

画像の見えを向上する方法の一つとして，濃淡のコントラストを強調する手法が提案されている．これらの強調手法は画像の処理領域に関して，局所的強調手法と大域的強調手法に大別される．局所的強調は局所的な視認性を向上させる反面，大域的コントラストを低下させる性質がある．逆に，大域的強調は大域的コントラストを増す反面，局所的な視認性を低下させる性質がある．本稿では，局所的強調手法をベースとして濃淡画像の領域分割処理を併用することにより，局所および大域双方のコントラストを強調可能な手法を提案する．すなわち，画像を適当な領域単位に分割し，画像全体と各分割領域との間，および各分割領域とその内部の各画素との間で局所的強調手法を階層的に適用してコントラスト強調を行う．これにより，大域的強調効果と局所的強調効果を同時に実現することが可能となる．局所的強調手法として逆S字型変換を採用し，従来最も強調効果の高い手法として知られるヒストグラム平坦化法と局所的ヒストグラム平坦化法を組み合わせた処理結果，および逆S字型変換を単独で適用した処理結果と実験的に比較した．その結果，提案手法によって，局所的にも大域的にも良好なコントラスト変換画像が得られることがわかった．

抄録全体を表示

Perceived image distortion of any image is strongly dependent on the local features, such as edge, flat and texture. In this paper, a new objective no-reference (NR) image quality evaluation model for JPEG coded images based on the local features and segmentation is presented. The local features information of the image such as edge, flat and texture area and also the blockiness, activity measures, and zero crossing rate within the block of the image are evaluated in this method. The results on two different image databases indicate that the model performs quite well over a wide range of image content and distortion levels.

抄録全体を表示

移動物体検出は動画像処理において重要な技術の一つであり，検出の結果がその後の処理に影響を与える．従って，動画像の中から安定に物体を抽出することが必要になる．背景差分法は，移動物体を検出する効果的な手法としてよく利用されている．しかし，この手法は処理アルゴリズムは単純であるが，濃度値の変化の影響を受けやすいという問題点がある．本稿では，正規化相関照合を利用した移動物体検出手法を提案する．正規化相関照合は，濃度値の変化にロバストであることから，動画像中の明るさの変化に対して安定な検出が実現できると考えられる．更に，照合結果の統合処理を行うことで，更に安定な検出を実現する．実験の結果，提案手法は明るさの変化の影響を受けずに，安定な移動物体検出が可能であった．

抄録全体を表示