Multi-resolution intrinsic texture geometry based local binary pattern for texture classification

In this paper, we propose a new hybrid Local Binary Pattern (LBP) based on Hessian matrix and Attractive Center-Symmetric LBP (ACS-LBP), called Hess-ACS-LBP. The Hessian matrix provides the directional derivative information of different texture regions, while ACS-LBP reveals the local texture features efficiently. To obtain the macro- and micro-structure textural changes, Hessian matrix is calculated in a multiscale schema. Multiscale Hessian matrix presents the intrinsic local geometry of the texture changes. The magnitude information of the Hessian matrix is used in the ACS-LBP method. A cross-scale joint coding strategy is used to construct Hess-ACS-LBP descriptor. Finally, histogram concatenation is carried out. Extensive experiments on eight texture databases of CUReT, USPTex, KTH-TIPS2b, MondialMarmi, OuTeX TC_00013, XU HR, ALOT and STex validate the efficiency of the proposed method. The proposed Hess-ACS-LBP method achieves about 20% improvement over the original LBP method and 1%-11% improvement over the other state-of-the-art hand-crafted LBP methods in terms of classification accuracy. Besides, the experimental results show that the proposed method achieves up to 32% better results than the state-of-the-art deep learning based methods. Especially, the performance of the proposed method on ALOT and STex datasets containing many classes is remarkable.