Orientation Pooling based on Sparse Representation for rotation invariant texture features extraction

Inspired by the characteristics of sparse representation and pooling in human visual system, Orientation Pooling based on Sparse Representation (OPSR) is proposed to extract sparse and rotation-invariant texture features. At first, we assume that the over-complete dictionary represents a population of neurons in the cerebral cortex, and each atom in it will respond to a stimulus with a specific orientation like the response of a simple cell in the visual cortex. Then, the atoms are rotated at several different angles and added to the dictionary. Thus, atoms in the extended dictionary can respond to stimuli at different orientations. The responses of each atom and its corresponding rotated ones are pooled to obtain rotation-invariant texture features, which simulates the invariant features obtained by pooling the responses to stimuli of different orientations in human visual system. The comparative experiments with several traditional methods on two texture databases are conducted. The results demonstrate that OPSR method can effectively extract texture features with stronger rotation invariance.