Extracting the feature of alignment in oral pathological images, in helpful for the pathological diagnosis and development of diagnosis guidelines. This paper propose two types of processing, basal cell counting by the distance transform and segmentation of the epithelium region using multiple threshold values which is calculated by a self-organizing map neural network. The counting results have shown a good agreement with these of professional histopathologists, and hence the proposed counting is sufficiently accurate. The segmentation of epithelium regions has been successfully developed with the aid of a set of multiple values for luma thresholding.
In the optical soldering inspection systems, the existence or nonexistence of soldering paste and unfirm solder are relatively easy to be detected. However, it is difficult to classify the amount, the shape and the state of the fillet of brazed joints. This study aims at a development of an automatic classification system of solder joints on printed circuit boards and an exploration of effective features for defective soldering detection by means of statistical image analysis of digital images. A number of features are selected from textural features to be tested for the classification into 5 typical classes of solder joints. Experiments have shown that they are effective.
The requirement of high quality image is increasing as the wide use of digital image. Edge enhancement is a important way to improve the quality of a image. As a important solution of edge enhancement, there is possibility of processing in frequence space. In this article, we try to use FFT(Fast Fourier transform) to convert a image into frequency space and used three types of filters to enhance the edge of the image. From the result we analyze the feature of three filters and comprehend the edge enhancement in frequency space.
In this report, we present a depth estimation method from a single still image using image structures. Stereo cameras are commonly used to estimate depth in a still image. However, they are very expensive and inconvenient for daily use. Although there have been some methods to estimate depth from a still image by considering scenes in it, these methods a learning process to obtain general scene features. It consumes very high computational costs and time. In our study, we developed a simple and efficient depth estimation method without learning process by using image segmentation and edge detection. Our method can successfully estimate depth from a single image especially if the scene in the image has a simple structure.
This paper presents a motion vector (MV) prediction method for efficient video coding. In the state-of-the-art video coding standard, H.264/MPEG-4 AVC, the predicted MV of the current macro block (MB) is calculated from neighboring MVs in previously determined MBs. Then, residual between the predicted and the accurate (block-matched) MVs is encoded. However, the residual would increase when the correlations among neighboring MVs are low. This paper discusses a new motion vector prediction method that eliminates unreliable reference MVs in neighboring MBs and calculates a predicted MV using adaptively weighted reliable reference MVs. The experimental result shows the capability to improve the coding performance.
We propose a human behavior model with the visual information that can be obtained from the environmental circumference and also a generation method of autonomous human behavior that is created based on the model. The visual information is constructed with two types: color and depth information that is generated by CG rendering at the view point of each person. We have applied our method to a 3D space that has many branches such as maze. As the result of the simulation, each human could recognize crossroads autonomously by using the visual information and could make the map of the environment. Finally, the person could reach the goal autonomously and we have confirmed that our method is effective to generate autonomous human behavior.