The Research About The Stereo Matching Algorithm Based On PatchMatch For High-Resolution Images

Stereo matching aims to find the vision correspondences in two or more images about the same scene.According to the correspondences in the images,the disparity map about the scene can be acquired.However,the high accuracy disparity map is still a great challenge for the researchers and engineers due to the factors such as image noise,occlusion region,depth discontinuity region and textureless region.A substantial of stereo matching algorithms have been proposed to obtain high accuracy disparity map.Traditional stereo matching algorithms can be classified into two strategies: feature-based stereo matching algorithm and region-based stereo matching algorithm.Recently,deep learning has been applied to solve many difficult problems in the field of computer vision and shows a surprising performance.Meanwhile,it is used in stereo matching.These three strategies have their own advantages and disadvantages.This dissertation studies related stereo matching algorithm for high-resolution images.For stereo matching,the running time of the algorithm to dealing with the matching image pairs is a vital standard for performance evaluation.Especially for high-resolution image pairs,the running time of some stereo matching algorithms may be several hours,several days or even more.This dissertation aims to reduce the running time and improve the accuracy of the disparity map for high-resolution image pairs.The innovations of this dissertation are as follows:(1)Four-moded Census transform was first introduced into stereo matching algorithm for high-resolution image pairs.Compared with non-parameter Census transform,four-moded Census transform can represent more image features and has good ability to overcome the environmental noise.Meanwhile,four-moded Census transform can reduce the time for image feature extraction and improve the accuracy of the disparity map in contrast to image feature extraction based three-color similarity.(2)The discrete disparity plane approximation strategy was proposed in this dissertation.This strategy is applied during the inference procedure of the disparity for high-resolution stereo matching algorithm to reduce the running time and improve the disparity map.(3)This dissertation also refined the original four-mode Census transform.The refined transform regards the intensity value of the central pixel as an important valueduring the transform process,which makes the refined four-mode Census transform better reflect the image feature about the image patch.The stereo matching algorithm based on PatchMatch for high resolution images can obtain more accurancy disparity maps when the refiend four-mode Census transform is introduced.