Abstract: Human visual system processes the received visual signals on different scales, however, the traditional stereo matching algorithms get the disparity map from the original image in the biggest scale, which will lead to high stereo matching error rate in the area of low-texture, texture-less area. Simulation of the human visual system in stereo matching on multiple scales can reduce error rate. A stereo matching method based on the Gaussian pyramid cross-scale transform was improved, by adding the Laplace Pyramid Transform to the original cross-scale framework and adding the weighted joint bilateral filter in the disparity refinement stage. The new cross-scale framework can get a better disparity map than the original method.

Key words: stereo matching, cross-scale, pyramid transform, weighted joint bilateral filter, disparity refinement