A Simple and Effective Deep Registration Algorithm for SAR Images

Synthetic aperture radar (SAR) image registration contributes significantly to various applications of SAR images. Conventional scale invariant feature transform (SIFT) alignment methods possess abysmal robustness for SAR image registration resulting from speckle noise in SAR images. Attributing to the excellent feature representation ability of convolutional neural networks (CNN), a simple and effective deep registration algorithm for SAR images is proposed in this paper. Specifically, we construct a CNN to extract the features of SAR images and to accomplish the feature matching, while outputting the transformation matrix for SAR image registration directly. Furthermore, a data generation strategy is adopted to solve the problem of insufficient training data during the training process. Experiment results on the real data verify the effectiveness of the proposed method.