Aiming at the problem of only estimating disparity from left and right stereo images, a novel method of stereo matching is proposed, which decomposes object images into a set of sub-bands by discrete wavelet transform, then modifies each sub-band image, and improves the matching quality of each sub-band image by implementing discrete inverse wavelet transform. On this basis, we estimate disparity of each modified image, and merge different sub-band disparity maps into a final disparity map using HVS model. Experimental results demonstrate that, compared with the existing algorithm, the proposed algorithm achieve a high-precision.