Traditional stereo matching algorithms are mostly based on the correspondence between two image pixels or partial blocks，finding the disparity map at a single scale． But these algorithms can not model the correspon-dence between low-texture and repeated texture regions，resulting in a limited accuracy of the obtained disparity map． Considering the human visual system processes the received visual signals on different scales，a cross-scale restart and random walk algorithm was proposed to improve the above problems． Firstly，the matching cost of the scene images was calculated． Then using super pixel segmentation for rapid initial aggregation，and using the re-start and random walk algorithm to optimize it globally． Finally，an effective fusion update of the matching cost was realized by adopting the cross-scale model，and then a disparity map of the scene image was obtained． The ex-perimental results on the Middlebury dataset show that，compared with the traditional cross-scale stereo matching algorithm，the proposed algorithm can effectively reduce the average mismatch rate of the scene image in all regions and non-occluse regions by 1 percentage and 3 percentage points respectively，and obtain high-precision disparity
map．