Referring to the problem of the less correlative intensity image registration, an improved approach based on Scale Invariant Features Transform (SIFT) is proposed. First, according to the performance of SIFT in three instances: affine transform, adding Gaussian noise and different intensities, SIFT vector values which was easily affected by different intensities were constrained through given threshold when extracting SIFT feature points. Then, the false matching points were deleted by similar quadrangle in the precise math process. Finally, least squares method was employed to find optimal solutions to affine transform equations. Experimental results show that the proposed approach has a better performance than original SIFT algorithm, and can achieve a high accuracy and moderate complexity level.