A SIFT matching algorithm based on adaptive contrast threshold

SIFT keypoints are rejected if the contrast is less than given threshold. The threshold has to be selected for different images manually and therefore cannot achieve automatic image matching. In order to improve the accuracy and automation of SIFT based image registration algorithms, a SIFT matching algorithm based on adaptive contrast threshold is proposed. Firstly, compute SIFT contrast threshold through normalized image entropy and then the threshold is used to reject low contrast SIFT keypoints. Secondly, the Nearest Neighbor algorithm is used to obtain initial matching set according to common matching threshold. Thirdly, the obvious mismatches are rejected using the histogram of main orientation difference of the initial matching set and then calculate the control parameter k of mathcing threshold. After that, matches with 8 smallest ratios are selected to solve the initial perspective transform parameters from the transformed image to the reference image and then calculate the maximum error σ from the 8 transformed keypoints to the 8 corresponding keypoints of the reference image. Finally, compute the error from the transformed keypoints to the corresponding keypoints of the reference image for all initial matches. Matches are rejected if the errors are larger than 3kσ, then the refined matching set is obtained. The experimental results show that the proposed method can effectively extract correct matches and is more robust to low contrast image matching. Moreover, the contrast threshold and the matching threshold are computed automatically for different images.