Due to the background noise and the detector noise, the centroid error of Shack-Hartmann wavefront sensor is relatively large. A method to filter the noise in sparse region is proposed. First, the sample image is generated by using the two-dimension Gauss model and the over-complete target dictionary is constructed. Then, the sub-image blocks of the test image are extracted successively and the corresponding coefficients are calculated with the constructed over-complete target dictionary. Since the coefficients between the noise and the target are large different, the difference by setting a threshold is used to subtract the target. Experimental results show that, the target can be well subtracted, while the centroid deviation and the RMS and PV of the centroid are smaller than using the method of subtracting threshold