An active vision system with associative memory can recognize overlapping patterns by shifting the fixation-point

In this paper, a new system for recognizing overlapping patterns by shifting the fixation-point using the associative memory of a neural network is proposed. This system deals with overlapping patterns of convex polygons by shifting the fixation-point at each vertex of a pattern. When the fixation-point on a pattern shifts, the appearance around each point changes accordingly. Thus all the appearances of the pattern to be recognized are memorized in the associative memory. The appearance around a fixation-point on overlapping patterns is fed to the associative memory of the neural network. The fixation-point is shifted from vertex to vertex. From the appearances around some fixation-points, one of the memorized appearances of the patterns which may compose overlapping patterns can be recalled. A scale invariant transformation is proposed. Each of the transformed pattern around each fixation-point is memorized in the associative memory and the overlapping patterns transformed around each fixation-point are fed to the associative memory as an input.