Abstract: In the paper, we present a method of interactive image segmentation with texture constraints in the framework of graph cut. Given an image, we first gather user-marked information to establish the color and texture prior models of the foreground and background. Then, we formulate an energy function composed of color, gradient and texture terms. At last, the foreground is extracted by minimizing the energy function using graph cut. In the energy function, the texture term describes the difference between the texture prior models and the texture descriptors of each pixel to be labeled. The foreground/background texture prior model is represented as histograms of Local Binary Patterns (LBP). Every pixel to be labeled in the image has a foreground and a background texture descriptor, which are obtained by a randomized texton-searching algorithm. The newly added texture term is effective to overcome the difficulty in locating real boundaries when dealing with textured foreground/background. Experimental results demonstrate that, with the same amount of user interaction, our method generates better results than traditional ones.