One-step method to take Integral Photography (IP) using a 2-dimensional concave lens array was proposed in a long time ago, but the characteristics of captured images has not been reported as far as we know. In this study, we theoretically showed the problems of this method, and some experiments were carried out to evaluate captured images using handmade 2-dimensional concave lens arrays. As a result, it was found that if capturing viewing zone is 30°(±15°) or less, relatively realistic 3-dimensional image can be taken with a spherical concave lens array, but the wider viewing angle increases the distortion of the elemental images, and decreases the reproducibility of ray angle. It suggests that a concave lens array with the curved surface which is optimized to fit the display device is needed for getting a wide viewing zone.

View full abstract

This paper discusses the solution ambiguity problem of linear 3D motion parameter estimation, especially concerning homography matrix decomposition. The solution ambiguity by the homograohy matrix decomposition is well known in the computer vision field. According to our simulations and consideration for this problem, we found that: (1) the ambiguity yields under some conditions, only one solution can be obtained with another condition; (2) the occurrence ratio of one solution and two solutions depends on the size of the feature region. We also give (3) a geometric explanation of this problem and a qualitative property of the cases which can obtain a unique estimated solution.

View full abstract

Image inpainting is very useful to restore or remove objects from digital images. Among several techniques to restore an image, exemplar-based inpainting is one of the most widely used. In instances in which large objects are removed, it dramatically outperforms earlier works in terms of both perceptual quality and computational efficiency. However, the exemplar-based approach has certain weaknesses such as high time cost and visual inconsistency in some cases with depth ambiguities. In this paper, we improve the exemplar-based approach by reducing the search space of exemplars. We obtain this reduction by using a combination of the wavelet transform of the image with an automatic computation of the search window. Numerical simulations show that the proposed approach considerably reduces the computational cost of original exemplar inpainting whilst keeps the quality of resulting images as good as in previous technique.

View full abstract

In graph-based pattern recognition, representative graph influences the performances of recognition and clustering. In this paper, we propose a learning method for generating a representative graph of a set of graphs by constructing graph unions with merging corresponding vertices and edges. Those corresponding vertices and edges are obtained using common vertices of a set. The proposed method includes extracting common vertices and correspondences of vertices. To show the validly of the proposed method, we applied the proposed method to pattern recognition problems with character graph database and graphs obtained from decorative character images.

View full abstract

This paper proposes a feature selection technique with genetic algorithm that reduces the risk of data leaks by reducing the visibility of gradient-based image features. Gradient-based image features, which are used in image classification, are capable of wide application and offer high classification accuracy. However, people can picture the original image in their minds easily from the features when used in high resolution because they are represent appearance. This creates privacy concerns when they are applied to face image recognition. To overcome this problem, we introduce a feature selection technique that uses the genetic algorithm to reduce the visibility of gradient-based image features without sacrificing the recognition rate significantly. To evaluate the performance of the proposed technology, we make an experimental feature selection system that incorporates gender classification software. An experiment shows that the proposed technology can well reduce the visibility of gradient-based image features without sacrificing the recognition rate.

View full abstract

This paper proposes a method to track the moving object in moving picture like digital movie using template matching. The template is converged to distinguishing characteristics of the object using the SN-ratio of matching rates. In order to stabilize the tracking, our method update the template image flexibly to adapt it to change of shape. The experimental results show that the proposed method had a better tracking performance than that of conventional methods.

View full abstract

Many methods of analyzing gait by foot pressure distribution time-sequential images have been studied. However, the measurement for training how to walk correctly is not yet established. In this paper we propose new gait parameters by which a walking instructor can teach directly how to move each part of the foot sole. These parameters are extracted by quantitatively evaluating the movement of heel, middle part and each finger. First, foot sole is divided into 10 parts by template matching. Then, partial load of 10 parts while walking is calculated. And finally, the following three kinds of parameters are detected, maximum load, phase at which partial load is maximum and duration of contact duration for each part of the foot sole. The experimental results show that the proposed method can be applied to training for a walking.

View full abstract

Signal-to-Noise Ratio (SNR) is widely used for objective measurement of image quality due to its simplicity and convenience. SNR of a color image is often calculated as an arithmetic mean of SNRs for each color component. However, if SNR is expressed using logarithmic scale, the SNR derived by this method does not coincide with the SNR obtained by arithmetic mean of power of signal error of each color component. In this paper, we analyze difference of these SNRs and point out how SNR should be calculated for multi-component images.

View full abstract