The genetic algorithm (GA) can search only one global solution. On the other hand, the immune system can search plural local solutions. Recently, the immune system is applied to the conventional genetic algorithm. One of them is the genetic algorithm with immune system. However it is possible only to search only one global solution caused by major improvements. This paper proposed a method to keep diversity of individuals and extract plural target image areas with free locations, inclination and scale that are similar to template image by using hill-climbing flight insect method. The “hill-climbing flight insect method” is an evolutional system introducing a viewpoint of both genetic algorithm and hill-climbing method, and can use discrete space like the genetic algorithm. The experimental results show that the proposed method can search more plural local solutions than conventional methods.

View full abstract

Mammography is a widely used screening tool for the early detection of breast cancers. On the other hand, screening by mammography imposes heavy burdens for medical doctors, and these burdens lead to misdiagnoses. In this paper, we propose a method for detecting mass regions in mammograms to reduce the burden. The purpose of this study is to improve sensitivity for mass detection reducing the number of false positives. We focus on accuracy of extracting mass region because its contour will be helpful information for medical doctors and automatic classification by computer aided diagnosis (CAD) system to classify whether it is benign or malignant. In the proposed method, we apply Mean Shift to segment into some regions based on density of image data. After the segmentation, we obtain concentration of gradient vectors using iris filter and detect regions of mass candidates. According to the field test with medical doctors, the proposed system is tested with 398 mammograms containing 193 masses. In the result of a performance test, the proposed method obtains a sensitivity of 81% at 5.0 false positives per image. In addition, we extract 75% masses, where the Area Overlap Measure (AOM) shows more than 60%.

View full abstract

Our method can create learning-friendly low-definition lecture video from a video captured by a high definition fixed camera, no specific equipment or training is needed. The proposed method automatically extracts the screen position, the speaker's dynamic position and the slide changes from the video of a lecture to produce a spatially-condensed low-resolution version of the video. We conduct user tests to evaluate the operation time and the quality of the videos produced by the proposed method and reach the following conclusions: (1) The proposed method can automatically generate about 640×480 pixel videos that are suitable for video distribution, then the full, high-definition videos (1920×1080 pixels) are subjected to non-linear reduction. (2) The edited videos well support learning. (3) All subjects could input the initial information needed for detecting positions of the speaker and the slide within just a few minutes.

View full abstract

At present, biometric authentications using bodily features are attracted, and they have been used in various fields because neither a password nor a key is needed. Especially, face authentication is required and many methods have been researched. Face authentication often uses a front face image, but has the disadvantage that treats too many features. This method performs authentication using the outline feature of a profile face image as a human physical character. The purpose of the paper is evaluating the possibility for the person authentication using the profile image. The method registers the outline feature from a captured profile image and calculates the matching similarity of pairs of pixels between the registered image and an input image. The experimental results showed that the authentication rate was about 89.4% and it was found that the profile face images contain useful features for the personal identification.

View full abstract