Abstract: Registration between SAR and optical images is time-consuming and has poor accuracy
when based on the scale-invariant feature transform (SIFT) algorithm. In this letter we propose a
novel method to solve this problem. First, we smooth SAR image by using bilateral filter (BF). BF is
also good at preserving edges in the image as opposed to Gaussian smoothing, which is used in the
original SIFT. Then, keypoints are detected in the Difference-of-Gaussian (DOG) scale space and
SIFT descriptors are generated. Next, we adopt the fast library for approximate nearest neighbors
(FLANN) algorithm which can search matching points fast in high-dimensional space. Last,
progressive sample consensus (PROSAC) algorithm is utilized to exclude false matches. Experimental
results show that our approach is significantly more accurate and much faster than the original SIFT.

Key words: synthetic aperture radar image, optical image, registration, scale-invariant feature
transform, fast library for approximate nearest neighbors