Applied connectionistic methods in computer vision to compare segmented images

Abstract

Two similarity measures to compare whole and parts of images are proposed. These measures consider the color, shape and texture properties of image segments as well as their relative positions mutually. After image segmentation MPEG-7 descriptors are computed for each segment. Now the information reduced images will be represented by labeled graphs. To compute the first similarity measure it is necessary to solve the Maximum Weight Clique Problem in an extended compatibility graph. This problem is solved by a connectionistic method. The second similarity measure needs to solve a special kind of Consistent Labeling Problem. This is done by a connectionistic method too. Both measures may be used to improve the similarity measurement of images for instance in image and video retrieval, object recognition, stereo and many other Computer Vision tasks.