High-performance SAR image registration algorithm using guided filter & ROEWA-based SIFT framework

Abstract

To address the performance degradation of SIFT-based SAR image registration algorithm caused by speckle noise and local deformation of SAR images, this paper presents a novel SIFT-framework algorithm for SAR image registration based on improved multi-scale space construction strategy and a novel local feature detection and descriptors. In our proposed algorithm, the multi-scale space construction is generated by Guided Filter because of its real-time and edge preserving. The feature detection section adopts Harris-Laplace combined with ROEWA, which is effective to suppress the false alarm on high-contrast areas of SAR image. Moreover, the feature description adopts the GLOH by ROEWA method, since the phase method of GLOH descriptor is robust to rotation. At last, we suggest using K-Nearest Neighbors (KNN) to speed up the search for quick rough match, and then using the random sample consensus algorithm (RANSAC) to remove false match points. Experimental results indicate that our proposed algorithm is real-time and produces better performance than SIFT-based methods on SAR image registration.