Unsupervised Texture Segmentation Using Total Variation Minimization
利用总变分最小化方法的无监督纹理图像分割

This paper is devoted itself to segmentation of texture images. Based on the theory of total variation minimization and the active contour image segmentation method, we proposed a simple linear model of texture images which regards a texture image as a sum of a photo prototype image and a texture sub-image. Using the total variation minimization method the simplified prototype image can be extracted from the origin image. A coarse border can be located by segmenting this simplified image. Based on the coarse border, a higher accuracy result can be obtained by taking the original image into account. We choose the geometric MDL active contour for image segmentation and applied AOS scheme for the numerical solution of the nonlinear diffusion equation of the total variation minimization. Our method is unsupervised. Experiments on both synthetic and natural texture images show that the method is effective.