Scale-adaptive local binary pattern for texture classification

Local binary pattern (LBP) has already been proved to be a powerful measure of image texture with fixed sampling scheme: all P neighbor pixels in a single-scale are usually sampled by using a fixed radius R. It can effectively address grayscale and rotation variations. However, the LBP method is sensitive to image noise and fails to achieve desirable performance for texture classification with significant scale changes. With the aim to deal with these disadvantages of LBP, a new method named scale-adaptive local binary pattern (SALBP) is proposed in this paper. The essence of our proposed SALBP method is to adaptively find a single but optimal scale from multiple scales for each radial direction in accordance with the characteristics of a local area. First, we select candidates of neighbors by the majority vote strategy for signs. Second, we determine final neighbor pixels by the maximal difference magnitude selection strategy for magnitudes of candidates. This procedure lets each neighbor pixel to adaptively obtain its own optimal scale from multiple scales, which refers to the adaptive optimal sampling radius. Therefore, scale invariance can be significantly improved. Extensive experiments on four public texture databases (i.e., Outex, CUReT, UIUC and Brodatz) demonstrate that the proposed SALBP method achieves significantly better results than representative LBP variants for texture classification tasks with a smaller feature dimension.