Abstract: To resolve the problem of large angle and large scale image registration, an improved approach based on scale invariant feature transform (SIFT) is proposed. The log-polar technique was applied to estimate the parameters of rotations scales between reference image and sensed image. The images were segmented into sub-blocks and six candidates of sub-blocks were extracted according to information entropy, where SIFT features and moment features were fused to form a new feature descriptor. The image registration result was calculated from the matching points which were obtained by combining the Euclidean distance and the algorithm of iterative procrustes. The experimental results show that the proposed method is fast with high-precision.