Research And Implementation Of Stereo Matching Algorithms Based On PatchMatch

Stereo matching is one of the popular topics of computer vision.It is realized by finding as many as possible matching point pairs in two images,that are taken from different views,and then used for three-dimensional reconstruction.However,in practical applications,there are lots of difficulties in stereo matching,like specular reflection,occlusion,discontinuity,camera distortion,and textureless region.Therefore,how to obtain high-precision disparity maps has always been the focus of the scholars.Traditional algorithms realize stereo matching by searching for matching pixels in a exhaustive manner on a pixel-by-pixel basis,which is inefficient.After the PatchMatch algorithm is proposed,researchers apply it to stereo matching,and find matching pixel of each pixel by matching the patch and propagating to improve the efficiency.In fact,some prior information like various features of the images can be utilized in the study of stereo matching.In this paper,the image feature is used as a matching factor to establish a feature detection PatchMatch algorithm that replaces the original random matching with feature matching.Through the experiments,this paper will prove that this feature matching based feature detection PatchMatch stereo matching algorithm is more efficient than the original random initialization based PatchMatch stereo matching algorithm.In order to further improve the accuracy of the algorithm,this paper applies a large-scale neighborhood window and the disparity post-processing to the feature detection PatchMatch algorithm to solve the common problems of stereo matching such as specular reflection and occlusion to some extent.Finally,to solve the problem of poor efficiency of calculations in the texture-free and texture-repetitive regions,a matching cost function based on image pyramid distance metric is proposed.The research content of this paper is as follows:(1)Pixel-level matching is achieved on the Middlebury dataset for four kinds of image features,and the matching results are evaluated using the standard disparity maps.The error rates and runtime of different features in stereo matching are summarized.(2)A stereo matching algorithm combining feature detection and PatchMatch algorithm is proposed.It can use the prior information to decrease the runtime and calculation of large windows,to solve the problems brought by the weak texture region or specular reflection region of images.(3)We apply Feature Detection PatchMatch to the proposed stereo matching algorithm based on the original PatchMatch,combined with post-processing,to solve the difficulty of calculating the accurate disparity in the occlusion regions.(4)A matching cost function based on image pyramid distance metric is proposed.In the same matching window and different scale images,this matching cost function can obtain a larger range of pixel information and structure information from the original image,especially the repeated texture or untextured image areas,which improves the calculation accuracy of the algorithm.