A New Image Segmentation Approach with Structure Tensor and Random Walk

This paper proposes a novel approach for image segmentation by first applying the structure tensor to random walk. The method firstly presents the definition of scale vector by analyzing the structure tensor, to calculate the weight between two graph-nodes. And then, apply the random walk algorithm to achieve segmentation. The nature of structure tensor helps the new weight retain more structure information of image, and that makes the segment more accurately. In addition, the paper also presents an anisotropy filter, replacing the Gaussian function, to average the structure tensor. The new filtering function makes structure tensor retain more information of complicated structure. In experiments conducted on various images, the algorithm shows a notable visual improvement.