An ICA Based Method for Texture Recognition

Abstract

The method proposed in this paper uses the Independent Component Analysis (ICA) for an application of unsupervised recognition of textures. The analysed texture is modelled by a weighted sum of almost statistically independent random signals that are extracted with FastICA algorithm. Each resulting signal is described by its negentropy, more precisely, by one of the approximations used by FastICA algorithm. The approximated negentropies are sorted into descending order and represented by a curve. The final step of the algorithm is the averaging of a certain number of such curves obtained from different zones of the texture. The resulting mean ”negentropy curve” displays a good discriminating power on the tested textures.