Robust and fast scale-invariance feature transform match of large-size multispectral image based on keypoint classification

Scale-invariance feature transform (SIFT)-based algorithms often suffer from false matches of keypoints while being utilized to register multispectral remote sensing images. It is even worse for the registration of large water areas due to the lack of keypoints and high similarity of textures in such regions. To tackle these problems, a robust and fast SIFT match algorithm for multispectral image registration based on keypoint classification is proposed. This algorithm establishes match candidates through classifying the keypoints based on their scale spaces, where the keypoint matching is restricted on the same scale group. To obtain an accurate scale group, the ground sampling distance of images to be registered is normalized, while the initial scale ratio is estimated, relying upon the seed matches. Furthermore, for refining the match candidates and improving the matching robustness on water areas of the image, the keypoints are classified into keypoints of the land region (L-KPs) and keypoints of the water area (W-KPs). The L-KPs of the match candidates with sparse region on the two-dimensional representation of the location offset are excluded by the iterative bivariate histogram, and then the average location offset of the correct land matches is employed to exclude the outliers of the candidates associated with the W-KPs. Meanwhile, to reduce the computational cost, the images to be registered are divided into corresponding overlapped blocks, providing parallel computing and local rectification. Experiments were conducted on large-size remote sensing images at different resolutions and the results demonstrate the effectiveness of the proposed approach.