Tensor Inhomogeneous Average Sparse Matrix Based Texture Extraction

Abstract

Texture extraction is considered as a basic but a very challenging work in a lot of computer vision fields. Yet texture is not precisely defined, which is difficult to be separated from edges. In this paper, a novel texture extraction algorithm was proposed. Nonlinear structure tensor was introduced to distinguish textures out from edges. And an 8-neighborhood tensor inhomogeneous average sparse matrix was presented to smooth the images. The smoothness weights are determined by the local anisotropy. By applying this inhomogeneous average sparse matrix to the input images, the textures are smoothed to the detail layer while the edges are remained in the original images. The effectiveness of our method was demonstrated by the comparison results with other existing generally acknowledged texture extraction algorithms. And the sparse matrix framework reduces the computational cost than the convolution frameworks.