Research On Image Matching Based On Graph Cut

Many computer vision problems can be described as energy minimization problems. For the energy minimization problem, traditional methods are sometimes use gradient decent and simulated annealing to solve and so on. However, gradient descent often falls into local minimum, and simulated annealing often has much longer convergence. Stereo matching problem in vision, the energy minimization based on graph cut has better robust and practical as the traditional optimizations, it can get a strong kind local minimum or global minimum. In this paper, the main work is focused on the theory of graph cut and the problem in the stereo matching. The main points are as follow:1. The research of the graph cut, it has introduced a special graph network-dual terminal graph. The graph has contained only source and sink points. By the research of the Ford-Fulkerson's algorithm, it gives the maximum flow and minimum cut solution process to the dual terminal graph. The graph has many terminals, it can construct virtual source point and sink point. Then, it transforms into the problem consists of dual terminal to deal with.2. For the research of energy function in the set of two variables F2 and the set of three variables F3. We analyze the regularity of the energy function. Energy function can be constructed a unified network on the two sets. Finally, we give the method based on graph cut for solving the energy function.3. To calculate the disparity of dense matching by graph cut, disparity in this matter can consider as label. Through the analysis of the data constraints and smooth constraints in the image matching, it must analyze the regularity of structure items. It can be constructed energy function and the graph corresponding to the energy function. Use a-expansion algorithm to solve the disparity of the corresponding points, it finally get the disparity of the image matching.