The conventional Fractal Dimension is scalar-valued and cannot contain sufficient image information for image analysis. For example, spatial information such as the pixels direction and relative arrangement of the pixels are not represented by the Fractal Dimension. Overcoming these defects, in this paper, a new type of fractal feature, which we call the Directional Fractal Dimension, is proposed and applied to texture analysis. The Directional Fractal Dimension is refinement of the Box Dimension, and is consisted of three Fractal Dimensions, considering directional cosine on which the pixels are placed. The Directional Fractal Dimension contains directional information for pixel placement. The recognition experiment is carried out for 10 texture images, and is obtained correct result 95.8%. Also, Directional Multi-Fractal Dimension is defined and the usefulness for image analysis is examined. The proposed new features are useful for image analysis and recognition.

View full abstract

The edge part of the image contains a lot of important information on the image character. Then, we examined the binary coding method by using the edge information named unevenness coefficient. The unevenness coefficient is computed by comparing the luminance value of each picture element with the average value of small area surrounding it. Even if the brightness of the screen changes, the density change can be precisely determined by the unevenness efficient. As a result, we found that our newly proposed processing concerning the shading image is more effective.

View full abstract

Stereo matching, the corresponding problem of stereo vision, requires much computational cost, especially highly accurate matching algorithm needs huge computational cost. Therefore realtime systems for stereo matching based on specific processors are developed. However, there is no system which enables both highly accurate matching and real time processing. In this paper, we propose a novel processor architecture based on variable window approach, in order to develop a specific processor which achieves highly accurate matching required for realtime systems using stereo vision. A processor based on the proposed architecture implemented on Xilinx Virtex-4 FPGA achieves real-time stereo matching at 70MHz when the resolution and the frame rate of input images are QVGA and 30fps, respectively.

View full abstract

A lossless coding of an error-diffused image by coding a multi-level image generated through inverse halftoning process using soft decision method similar to Viterbi decoding is proposed and studied. Through the study, block-type error diffusion is introduced, and entropy-based method and prediction-error-based method for inverse halftoning are compared. The result shows that the decoded image by proposed methods has about 30-40% less entropy than that of the original multi-level image. Finally it was expected that the entropy of decoded multi-level image will be even less than that of bi-level image in higher resolution case.

View full abstract

The lossy coding scheme of error diffused images is an interesting topic but has not been studied well yet. In this paper we study two different categories. First one is to use multi-level lossy image coding methods such as JPEG, JPEG-LS and JPEG2000 including reduction and enlargement processing. The other category is to send the number of black pixels in plural pixels, and we studied the simplest case of sending three levels in two pixels. Among the former category, we found that JPEG2000 achieves best, and compared the SNR of inversed halftone images at the same bit rate with the latter. Then it is shown that the first category works better for the low-frequency dominated images and the second one works better for the high-frequency dominated images.

View full abstract

Stereoscopic displays are required in various fields and the researches have been done briskly. In these researches, the three-dimensional display without glasses attracted attention in recent years. The parallax barrier method and the lenticular lens method are the main methods of three-dimensional display without glasses. These methods set the sub-pixels of each parallax image as vertical stripes by turns according to a slit or a lenticular lens, to display a stereoscopic image. However, when extending it into a multi-view system by setting many parallax images as vertical stripes, only horizontal resolution falls, so the quality of the stereoscopic image deteriorates remarkably. In this paper, a parallax barrier method with the good balance of horizontal / vertical resolution is proposed. This three-dimensional display balanced horizontal / vertical resolution by using a vertical barrier. Because horizontal / vertical resolution has good balance in resolution, when extended into a multi-view system, deterioration of the quality of a stereoscopic image is limited, so more multi-eye system is enabled compared with the conventional technology.

View full abstract

A lot of concrete structures such as dams, tunnels and highways have been built since period of high economic growth in Japan because those are lasting for years and are modeled easily by form mold. In consequence, concrete structures are essential in today's society. However, concrete structures need repairing because aged deterioration and natural disasters cause the structures crack and punctate depressions. In this paper we propose a method of automatic crack detection from concrete surface image by using mathematical morphology in order to support repairs. Our method generates a seed image and a paint image by morphology with five kinds of structuring elements. We extract crack region using those two images and verify effectiveness of our method with some experiments.

View full abstract

In this paper, we discuss about the modeling of cooking that correspond to real world and compose of status of objects and foods in a kitchen, and about the Cooking Navigation System synchronizing with user's action. Because the relationship between action and sensor events was not researched enough until now, it's not clear how useful the recognition of cooking action is. To confirm the effectiveness of action recognition using task model, we analyzed the recipe text and got the result that about 80% action can be recognized using information of object's location and cooking equipment's state. Then we implemented the cooking navigation system using cooking task model and experimented with the system. And we confirmed that it is possible to realize the cooking support system which synchronizes with user's action in ordinary cooking environment.

View full abstract

[in Japanese]

View full abstract