[in Japanese]

View full abstract

A visualization system called “Multi-motion” has been developed. It shows a gymnast's form and locus in the vault. It extracts player's image area quickly and automatically, and displays image of their form sequentially from the jump to the finish, overlaying the images on a scene like a stroboscopic camera, even in the case where the background has a complex pattern including moving objects such as walking spectators. To achieve this, a new extraction method has been developed. This method uses motion analysis based on the facts that “A player is bigger and moving faster, and spectators in the background are smaller and moving slower.” The system was used in the live broadcast of the World Gymnastics Championship and it gained favor in the public and gymnastic community.

View full abstract

A scalable image coding method is presented. Spatial scalability is implemented by using a multiresolution representation produced by a wavelet expansion. Quality scalability is implemented by using a multiprecision representation produced by multistage zero run-length coding after nonlinear quantization. The data rate is controlled by threshold values and/or spatial resolution in the wavelet expansion. The coding method owes its fine scalability and higher compression to multiresolution decomposition, nonlinear quantization, multistage zero run-length coding and arithmetic coding. In addition to the simplicity in coding, edges and outlines are well preserved without blocking noise at low bit rates. A few examples are also given.

View full abstract

Function regression can be viewed as template matching in an augmented space spanned by independent variables and function values. This formulation of function regression enables us to reject out-lier data and to preserve discontinuities in functions. In this paper, such a function regression method based on fuzzy Hough transforms is presented. The implementation of this approach by using neural networks is illustrated, and a supervised learning algorithm based on function interpolation of sparse data is proposed. The present method is used in image smoothing, segmentation by clustering of image pixels, and is also used in random dot stereo vision including transparent patterns.

View full abstract

In this paper we propose a support system for the visually impaired. This system enables the visually impaired to do daily activities without the direct support of others. It comprises a digital still camera and a pair of portable telephones. When a visually impaired person needs help, he takes a photograph of the object or the scene around him with the digital still camera. This photograph can be treated as digital image data. The image data are compressed and transmitted to a support person by the portable telephones. The support person will be able to understand the situation or trouble of the visually impaired person from the images received. And then the support person can give the proper advice or instructions to the visually impaired person. We investigated the circumstances where a visually impaired person needs the support of others. And we examine whether the system could be used or not in those circumstances. As a result, we confirm that this system is useful for navigation while walking, finding a lost article, and assisting their daily activities.

View full abstract

This paper analyzes the ways eye movements are affected by some factors : the impairment magnitudes, the picture contents, different kinds of assessors, subjective evaluation methods, different types of impairment. Hi-Vision picture quality, degraded by the bandwidth limitation and the addition of the white noise, is evaluated using the SSQS (Single-Stimulus Quality Scale) method and the EBU method (officially, Double-Stimulus Impairment Scale method). First, it is shown that the subjective evaluation method affects the absolute objective measure (total fixation duration, number of fixation points, moving distance between fixation points), the ratio of total fixation duration to judgment duration, and the objective measure concerning the moving distance. Next, the ratio of fixation duration in each region, divided by the contents of the test pictures, to total fixation durations is analyzed, and the validity of the description in the ITU-R Rec. 500-6, such that the SSQS method is suitable for evaluating the overall picture quality and the EBU method is suitable for evaluating small impairments, is confirmed. Finally, examined is whether the stability of evaluation is connected with any objective measure regarding the eye movements, the fixation duration ratio in each region divided by the picture contents, and the fixation duration distribution. Consequently, it is shown that the assessor's ability to discriminate picture quality is connected with the fixation duration ratio in each region divided by the picture contents and the fixation duration distribution.

View full abstract

Integral Photography (IP) is one ideal way to create three-dimensional (3 D) autostereoscopic images, because a 3 D autostereoscopic image can be seen from desired viewpoint without the aid of special viewing glasses. We have already proposed the IP method using a micro lens array and an HDTV camera to offer 3 D autostereoscopic images of moving objects in real time. But there are two problems yet to be resolved. One is pseudoscopic images that show reversed depth representation. The other is interference between the element images that constitute a 3 D autostereoscopic image. In this paper, we describe a new IP method with a lens array consisting of gradient-index lenses designed to conquer these two problems. Experimental results indicating the advantages of this method are shown. These results suggest the possibility of using gradient-index lenses for the IP method.

View full abstract

As the first step for developing the system that generates facial caricatures automatically, we developed a method of extracting the facial region and the irises. The facial region is extracted by a thresholding procedure in which the threshold is selected based on the degree of the coincidence between the outermost edges and the region boundary. Then, another type of thresholding is performed in the facial region, and the boundary of the binary image is detected. Next, using a Hough transform, a number of small circles are detected as candidates for the irises. On the basis of the geometrical relation between the left and right irises, a pair of circles are finally selected as irises. We experimented with 30 face images. The facial regions were successfully extracted for all of images and irises were successfully recognized for 29 images. It was found that the unsuccessful case could be recovered by a verification process.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract