Research Of Stereo Matching Algorithm Based On Image Segmentation

Stereo matching algorithm is to obtain the depth information of each pixel through two or more images,which is one of the core issues in stereo vision.The research of stereo matching has become an important research content of many computer vision researchers with the development of computer technology.At the same time,many good algorithms have been applied to the actual scene,including 3D reconstruction,target tracking and image refocusing.Considering the parallax continuity constraint of the stereo matching algorithm,except for the image boundary,the parallax of the central pixel and the parallax of the surrounding pixels do not change stepwise.According to this constraint,the image can be segmented and assumed within the segmentation region that the pixel disparity is continuous.Therefore,the paper combines the image segmentation algorithm with the stereo matching algorithm.The image segmentation algorithm has two advantages.On the one hand,it guarantees the disparity continuous,on the other hand,it starts from the superpixel level,reducing the time complexity of the algorithm.The research work in this paper is to fuse the image segmentation algorithm into two algorithms: non-local stereo matching algorithm and global stereo matching algorithm.The specific implementation is as follows:(1)A stereoscopic matching algorithm based on density-based image segmentation and non-local cost aggregation is proposed.Firstly,the basic principle and specific cost aggregation process of non-local cost aggregated stereo matching algorithm model are introduced.Then the mainstream image segmentation algorithm SLIC algorithm is studied.The shortcomings of the algorithm are analyzed.A density-based image segmentation algorithm is proposed.Finally,the density-based image segmentation algorithm is merged into the non-local cost aggregation model to achieve stereo matching.In the non-local stereo matching algorithm based on density image segmentation,the process of establishing the minimum spanning tree and the weighting formula of the edge in the process of cost aggregation are changed with the different stages of the spanning tree.Experiments show that the proposed algorithm achieves significant improvement in reducing mismatch rate and improving matching accuracy.(2)A density-based image segmentation algorithm and a global stereo matching algorithm based on PatchMatch-BP algorithm are proposed.In order to achieve the matching precision at sub-pixel level and improve the matching effect in weak texture and occlusion region,and reduce the time complexity of the algorithm,a PatchMatch-BP global stereo matching algorithm based on density segmentation is proposed.Firstly,the image is segmented by density-based image segmentation algorithm,and the same random initial parallax is given to each pixel of the segmentation region.Then,the neighborhood propagation and random propagation algorithm of PatchMatch algorithm are used to propagate at the superpixel precision level.Then,the pixel-level precision is used topropagate the confidence inside the superpixel,and finally the parallax with the smallest energy function is obtained to obtain the matching result.Experiments show that the matching effect of the proposed algorithm on sub-pixel precision is significantly improved compared with the existing algorithms in terms of improving time complexity and reducing false matching rate.