Supervised texture segmentation using localized dictionary based data modelling

In this paper, we propose a supervised algorithm for texture segmentation that uses sparsity based localized data modelling. The problem addressed is to segment a given test image whose constituent textures are known a priori. Overlapping patches are extracted from the texture. Each texture is modelled by learning the patterns of the patches that constitutes training data set. For each set of training data, a set of dictionaries are learnt, contrary to the conventional practice of one dictionary for all the patches of a texture. Each dictionary is learnt to capture the local pattern in the texture data. Texture is modelled by two level pattern learning. At the first level, clustering is used to learn the macro variations in the data pattern. Subsequently, data pattern in every cluster is modelled by a sparsity based subspace learning. These are subspaces where actual texture data lie. The set of subspaces are captured by learning a dictionary. The advantage of this approach is the accurate modelling of local data patterns which a conventional single dictionary is incapable of. Simulation results validate the proposed claim by achieving higher segmentation accuracy.