Texture Classification Using Local Derivative Binary Pattern

Abstract

Traditional Local Binary Pattern (LBP) descriptor is too coarse to preserve the more detailed local structural information for texture classification. In order to overcome this demerit, we present a novel rotation invariant approach, which is called local derivative binary pattern (LDBP) for texture classification. The local derivative binary pattern quantizes the intensity differences between the central pixel and its neighbors of the local regions in an adaptive way. In order to make the descriptor invariant to rotation, the differences are then ordered in a non-descending way and the ordinal information is encoded with binary codes, named LDBP_D. In addition, the gray level of the central pixel is quantized to further improve the discriminative ability, named LDBP_C. Both LDBP_D and LDBP_ C are combined jointly to represent the distributions of the texture. Also, our descriptor is robust to illumination variations and other geometric transformations. The experimental results obtained from three representative texture databases (Brodatz, Outex, and CUReT) show that our descriptor can achieve impressive classification accuracy compared with the existing state-of-the-art descriptors.