Neural Network Approaches for Attractive Area Extraction from Video Images

On-line machine vision algorithms are potentially applicable for various systems, including robot vision, security and guarding systems, clinical image diagnostic systems, and automatic video editting systems. However, in processing tremendous amount of image data in a stream of video signal, software burden required for image understanding sometimes restricts the realization of the systems. In our biological system, we first extract some essential scenes out of continuously in-coming stream of visual data given to the eyes both in temporarily and spatially, before through analysis. Furthermore, in understanding an image, our attention is focused on some narrow area of the image, e.g., the face of a person. We realized the above functions in neural network approaches. In finding out the important sub-area,we used three indices; color and brightness index, space frequency index, and symmetry index. Using hierarchically arranged BP networks, we realized such image processing with high speed. At the training stage, we gave the correct answers, obtained from human subjects through visuo-psychological experiments. After finishing with the training, the network system revealed reasonable results even to untrained images. This system will be helpful for data compression required for high-speed analysis in automatic image recognition systems as well as for editting video programs.