Selforganization by evolution strategy in visual systems

Abstract

Pattern recognition usually distinguishes two fields in its systems, feature detection and classification. There is a certain similarity to vision in man where feature detection partly takes place in the retina while the process of assigning the features to the proper concepts is supposed to be located in the cortex. With respect to a certain set of pictures one may ask whether the features being detected are the most advantageous ones for the solution of the classification problem. Within the system of parallel distributed pattern recognition being developed at the Department for Bionics and Evolution Techniques a certain feature is given by a local filter which is described by its structure and by some parameters contained in the structure. The locally detected signals are added and the result is a global value of that particular feature. If the structure of the local filter is well defined there will be no problems with the determination of parameters by applying evolution strategy.

In this paper, however, we will show how the structure of a local filter can be developed by means of evolutionary self-organisation. The optimization of structure is superimposed on the optimization of parameters. The task of the local filter, treated as an example, is counting the coherent areas in binary pictures. The self-organization by means of evolution strategy is carried out with 24 pictures and the result will be tested on 12 different pictures.