Abstract

This article describes three bio-inspired Texture Segmentation Architectures that are based on the use of Joint Spatial/Frequency analysis methods. In all these architectures the bank of oriented filters is automatically generated using adaptive-subspace self-organizing maps. The automatic generation of the filters overcomes some drawbacks of similar architectures, such as the large size of the filter bank and the necessity of a priori knowledge to determine the filters' parameters. Taking as starting point the ASSOM (Adaptive-Subspace SOM) proposed by Kohonen, three growing self-organizing networks based on adaptive-subspace are proposed. The advantage of this new kind of adaptive-subspace networks with respect to ASSOM is that they overcome problems like the a priori information necessary to choose a suitable network size (the number of Biters) and topology in advance.