Local binary circumferential and radial derivative pattern for texture classification

Motivated by gradients polar coordinate representation and its discrete approximation.Investigating the theoretical scheme of local differential approximation to build feature.Circumferential and radial derivatives pattern to capture tangential and radial information of derivatives, respectively.New descriptor is built based on radial plus tangential components and derivative information.Our approach produces high classification performance on several texture databases. Building discriminative and robust texture representation to deal with the changes of texture appearance is a fundamental issue in texture classification. The Local Binary Pattern (LBP) and its variants gain a lot of attention during the past decade and achieve great success in texture description. However, the current existing LBP-based features which treat LBP as local differential or orientation gradient operator, exploited local orientation pattern or anisotropic structure information separately. In this paper, we investigate the theoretical scheme of local differential approximation on the polar coordinate system in order to build a new LBP-based descriptor which better takes into account both radial plus tangential components and derivative information. First, we present an operator called circumferential derivative (CD) based on the tangential information with different order of derivatives. Then, we present an operator called radial derivative (RD) based on the radial information with different order of derivatives. Both extract complementary information locally around a central pixel. A new descriptor, the local binary circumferential and radial derivative pattern (CRDP) is constructed to fuse both local circumferential and radial derivative features based on different orders as well as a global feature based on global difference (GD) of central pixels intensity. Extensive experiments on Outex, CUReT, KTH-TIPS and KTH-TIPS2-a texture datasets indicate that the proposed CRDP descriptor is discriminative and robust. The results obtained by the proposed CRDP descriptor outperforms more than twenty recent LBP-based state-of-the-art methods, including the best reported results in the literature for aforementioned texture datasets to the best of our knowledge.