Image Texture Synthesis And Shrink Using Graph Cuts

Graph Cuts has been widely applied in fields of network flow analysis, computer vision, computer graphics, and so on. In recent years, Graph Cuts has been used for image processing to solve the seam optimization problem, which has become the new hot spot in image processing research. In this paper, based on Graph Cuts, image texture synthesis, image and video shrink are studied.An image texture synthesis method based on Graph Cuts and Ford-Fulkerson Algorithm is studied. Firstly, the texture synthesis problem is transformed into the problem of energy minimization. Secondly, an undirected flow network model is constructed according to the differences of pixels in the texture overlap region. Boundary of texture blocks can be found by Graph Cuts to realize texture synthesis. The new synthesized texture block is overlapped on adjacent blocks to calculate a new seam, so boundaries of texture blocks are optimized.The problem of image shrink is studied using Graph Cuts and Pre-flow Push Algorithm. A directed flow network is constructed according to the gradient of image pixels, and a seam with minimal energy is found by Pre-flow Push Algorithm. The pixels on the seam are deleted to realize image shrink. When the image with salient structure features is shrinked, it usually tends to appear a serious distortion. The capacity of arc in corresponding energy map is evaluated with the sum of absolute value of pixel 8-adjacent direction gradients. So, the surrounding region of pixel is fully considered, and good results are achieved.Video shrink is different from image shrink. Considering the temporal coherence of video sub-frames, three methods are studied in this paper. In the method of weighting the temporal gradient, the temporal gradient of each pixel in frame is weighted and projected to the corresponding arc capacity in energy map. A seam with minimal energy is found by Mincut-Maxflow and one horizontal or vertical seam is deleted to shrink one row or column each time, then the operation carries out alternately. In the method of corresponding arc weighted, the capacity of arc in energy map of adjacent frame is weighted and projected to the capacity of corresponding arc in energy map of previous frame. The corresponding arc is that locates at the same absolute position of two adjacent energy map. In the method of spatiotemporal complex processing, backward and diagonal arcs which are evaluated with infinite value are added on the same absolute position of two adjacent frames to construct a spatiotemporal energy cube, and a curved surface with minimal energy is achieved by Graph Cuts, which is monotone and continuous in temporal. The intersection between curved surface and adjacent frames respectively brings an optimal seam. The seam in previous frame is deleted to shrink its size. Processing for each frame, its spatiotemporal energy cube is constructed combining its next frame. By experimental comparison, the later two methods can effectively control the direction of pixels deletion and performs outstanding on keeping the image features.