Self-organizing scaling filters for image segmentation

The winner-take-all (WTA) learning mechanism and the self-organizing learning rule are shown to be suitable for developing overlapping circular filters of local Gaussian character. The range of filter sizes can scale with the size of the input samples and show the filter size clusters with gaps between. Experiments on identical filters suggest that a larger variety of input patterns will not modify the performance. Experiments on Hebbian and anti-Hebbian (HAH) networks show similar, but somewhat inferior results. To reach identical performance results suggest that competition in HAH should be increased in a smooth fashion, just like the Kohonen feature maps. In all cases (WTA and HAH), filters are distributed evenly over the input space. The self-organizing network described in this work can perform parallel image segmentation task.