Multiclass Color-Texture Image Segmentation Based on Random Walks Framework Integrating Compact Texture Information

In this paper, we propose an interactive multiclass color-texture image segmentation method. A new feature descriptor is designed by using the covariance matrices of coordinates, color with compact texture information and then integrated into random walks method to obtain the segmentation result. In this paper, we use multiscale nonlinear structure tensor (MSNST) to describe the texture feature of an image. Since the MSNST matrices set have different feature structures from color and coordinate vector, they cannot be used to construct covariance matrices directly. To address this problem and obtain the compact texture information simultaneously, we use the Isometric Mapping (Isomap) dimensionality reduction techniques for each scale of MSNST in tensor space. Experiments using synthesis texture images and real natural scene images demonstrate the superior performance of our proposed method.