Abstract: Proposes a new approach to image segmentation through combining the structure tensor with random walk algorithm. Taking advantages of the structure tensor, the weights between pixels are computed in terms of scale vector, then the image is segmented by the random walk algorithm. Because the weight used is in terms of scale vector, it can characterize the structure information so as to make the segmentation more accurately. In addition, an adaptively anisotropy filter function is proposed instead of the Gaussian function to smoothen the structure tensor, thus enabling the smoothened structure tensor to keep more information on image's complicated structure. The final experiments demonstrated that the proposed algorithm is more accurate and stable in image segmentation.