Automatic Image Classification by a Granular Computing Approach

In this paper we propose an image classification system able to solve automatically a large set of problem instances by a granular computing approach. By means of a watershed segmentation algorithm, each image is decomposed into a set of segments (information granules), characterized by suited color, texture and shape features (segment signature). Successively, images are represented by a symbolic graph, where each node stores the segment signature and edges retain the information about the mutual spatial relations between segments. The induction engine is based on a parametric dissimilarity measure between graphs. A heuristic search procedure based on a genetic algorithm is able to find automatically both the segmentation parameters and the dissimilarity measure parameters, and hence the relevant features to the classification problem at hand. System performances have been measured on the basis of an image classification problem repository which has been specifically created to this aim.