LSP: Local similarity pattern, a new approach for rotation invariant noisy texture analysis

Characterization of two-dimensional textures has many potential applications such as remote sensing, content base image retrieval, image segmentation, etc. In real world, noise has a disturbing effect in the analysis of images and textures. In this paper, a new rotation invariant texture descriptor, LSP (Local Similarity Pattern) is proposed to characterize the local contrast information based on the similarity or dissimilarity of adjacent pixels into a one- dimensional LSP histogram. The aligned histogram could be used as a feature vector to describe the related texture. Experimental results show that the proposed LSP operator can achieve significant improvement in the classification of textures in spite of their embedded noise. Especially, increasing the noise has a few effects on the performance of this method.