Abstract: It is difficult to extract the corresponding features from the multi-source images in automatic registration between them. Aiming to this problem, a new registration method based on the nonsubsampled contourlet transform (NSCT) and morphological shrink operator (MSO) was proposed. The feature extraction method based on NSCT_MSO can reduce the differences in angle and scale, and extract key structural feature points in multi-scale and multi-directional space. The feature matching method based on normalized mutual information computed from the low frequency band and the triangular consistency inspect method can extract a considerable number of corresponding feature points with even distribution, which ensure a high accuracy for the registration between multi-source images. The performance of the proposed algorithm was demonstrated and validated by experiments on SPOT-5(P) and ASTER images with considerable differences in angle and scale. The experimental results indicate that many corresponding feature points with even distribution can be obtained with the new algorithm and the accuracy of registration model is close to 1 pixel. The research can provide a basis for image fusion and object recognition.