Graph-Cut-Based Stereo Matching Algorithm Research Using Image Segmentation

In vision detection technology, the images matching is a key step. It is that given two images of the same scene, searching the corresponding relations between the pixels of the same scene point project to images. In stereovision, images are obtained from different visual angles at the same time, matching give depth clew. In moving image sequence, images are obtained from different time, then matching give motion clew. Compared with traditional algorithms (such as simulated annealing algorithm, M-estimate algorithm and so on), graph cuts algorithm not only has a high precision as a whole, but also more precise than other algorithms in discontinuity areas and low texture areas. Even if some algorithms'precision (for example simulated annealing algorithm) is close to the graph cuts algorithm's, the graph cuts algorithm converge quicker in the optimizing process.This paper presents a graph-cut-based stereo matching algorithm using image segmentation on the basis of in-depth study on the images matching algorithms. First, the reference image is divided into segments. Then modeling disparity inside a segment by a planar equation. A set of disparity layers is extracted from initial disparity segments in a clustering process. Each segment is then assigned to exactly one of those disparity layers. A global energy function is constructed. It measures the quality of assignments on the pixel and segment levels. Then the problem of matching could be transformed into that of energy function minimization. Using the knowledge of network flow, a network is constructed such that the energies flows theory, and hence making energies relate with capacities of network's cuts. Then using the images matching algorithm based on graph cuts theory, through the maximum flow-minimum cut theorem (a network's maximum flow equal to its minimum cut's capacity) obtaining the minimization of the energy function, thereby achieve the matching of images.Experiments demonstrate that the proposed algorithm produces good-quality results, especially in regions of low texture and close to disparity boundaries. The expensive computing cost for traditional global algorithms can also be resolved. Thereby, the algorithm can meet both demands for high resolution and real time.