Texture segmentation using image decomposition and local self-similarity of different features

Abstract

In this paper, we propose a new two channels feature space active contours model for texture segmentation by using image decomposition and local self-similarity descriptor of textures. The piece-wise smooth component of image decomposition is regarded as one channel of feature space for segmentation. Defined as a symmetry matrix and a kind of features fusion tool, the local self-similarity descriptor SSM captures the internal geometric layout of local repetitive pattern regions and is computed on different features of textures including oscillatory component of image decomposition, phase congruency and log-Gabor filters responses. A distance map dSSM can measure the similarities between the descriptor of template and the local windows surrounding every pixel on the texture image. And then dSSM is set as another channel of feature space for segmentation. Based on the space, texture segmentation is performed by using active contours and level set technology. In addition, the accuracy of texture boundary localization and the template searching inside initial contour are also concerned in this paper. Compared with some recent approaches, our method is more convincing and works well for synthetic textured images and ones in the real world.