Self-Organizing Fusion Neural Networks

Abstract

This paper presents a self-organizing fusion neural network (SOFNN) effective in performing fast clustering and segmentation. Based on a counteracting learning scheme, SOFNN employs two parameters that together control the training in a counteracting manner to obviate problems of over-segmentation and under-segmentation. In particular, a simultaneous region-based updating strategy is adopted to facilitate an interesting fusion effect useful for identifying regions comprising an object in a self-organizing way. To achieve reliable merging, a dynamic merging criterion based on both intra-regional and inter-regional local statistics is used. Such extension in adjacency not only helps achieve more accurate segmentation results, but also improves input noise tolerance. Through iterating the three phases of simultaneous updating, self-organizing fusion, and extended merging, the training process converges without manual intervention, thereby conveniently obviating the need of pre-specifying the terminating number of objects. Unlike existing methods that sequentially merge regions, all regions in SOFNN can be processed in parallel fashion, thus providing great potentiality for a fully parallel hardware implementation.