Robust Image Description with Weighted and Adaptive Local Binary Pattern Features

The Local Binary Pattern (LBP) is an effective image descriptor. However, this descriptor has limitations in some of challenging issues in texture analysis, such as invariance to scaling, rotation, viewpoint variations and non-rigid deformations. In order to overcome these demerits of LBP, the paper proposed a weighted and adaptive LBP-based texture descriptor. Adaptive definition of circular neighboring set in LBP descriptor is very effective to achieve scale invariance features [1]. In our proposed method, both circular neighboring radius and orientation of sampling in LBP descriptor are defined in an adaptive manner. We used the radius of blob-like structures to determine the radius of sampling, similar to [1]. Definition of LBP operator with respect to dominant orientation of each pixel can guarantee the rotation invariance of LBP features. The original LBP operator discards the magnitude information of the difference between the center and the neighbor gray values in a local neighborhood. Therefore, the paper also proposes weighted LBP features as a simple and efficient method to incorporate this information into the LBP histograms. We report extensive experiments comparing the proposed method to seven LBP-based descriptors, in texture retrieval and classification on two databases: Brodatz and UIUC. The experimental results show that the proposed Weighted-, Rotation- and Scale-Invariant Local Binary Pattern (WRSI_LBP) operator can achieve significant improvement in texture retrieval and classification over other LBP-based methods.