This paper presents an efficient binocular correspondence algorithm based ondyadic wavelet transform. Compared with the conventional algorithm, this methodachieves higher accuracy and speed (improved by about one order of magnitude). Thestereo images are first d6composed on the dyadic wavelet bases. By analyzing the transform coefficients, the authors get an efficient method to remove noises and extract edges.Some basic constraints for matching, e. g. the one which requires that the edge directionsbe nearly the same, are naturally expressed by the coefficients. More importantly, theycan deduce four useful constraints for selecting the potential matches which they call thesmooth component gradient constraints and the smooth component Laplacian constraints.They have strong disambiguation power and greatly remove those potential matches thatwould cause mismatches, and thus greatly improve the accuracy and speed for matching.