High-Performance SAR Image Matching Using Improved SIFT Framework Based on Rolling Guidance Filter and ROEWA-Powered Feature

Severe speckle noise and complex local deformation in synthetic aperture radar (SAR) images degrade the performance of common scale-invariant feature transform (SIFT) based SAR image matching algorithms. To address this issue, this paper presents a new SIFT framework based algorithm by designing a novel nonlinear multiscale space construction strategy and a new local feature detector and descriptor. In the proposed algorithm, the multiscale images are generated by embedding a rolling guidance filter into the procedure of scale space construction, which can achieve better results in edge-preserving despeckling than other strategies. The improved nonlinear multiscale space construction strategy lays a solid foundation for high-quantity feature point detection with a low false ratio. For feature detection, a new feature detector, Harris–Laplace technique combined with the ratio of exponential weighted average, Harris–Laplace ratio of exponential weighted average (ROEWA), is proposed to effectively suppress the false feature points in SAR images. Moreover, a gradient location and orientation histogram powered by ROEWA is designed to obtain feature descriptors, which is robust to local deformations. Finally, the K-nearest neighbor approach is used to speed up the search for initial matching; the spatially consistent random sample consensus is adopted to remove false match points (outliers). The experimental results using simulated and real SAR images demonstrate that the proposed algorithm can achieve much better performance in terms of the matching accuracy and inlier ratio than other methods.