Registration between SAR and optical images based on dual-channel network

Abstract

Traditional image registration methods rely on handcrafted feature extraction and matching strategies, which are inherently subjective and fail to fully exploit data-driven mechanisms. To address these limitations, we propose a novel dual-channel network for optical and synthetic aperture radar image registration. Our method employs two parallel convolutional neural networks for keypoint detection: RF-Net extracts shallow features with progressively expanding receptive fields, while ResNet captures deep global semantic features. The deep network channel provides high-level features that are fused with low-level features from the shallow channel, enabling robust keypoint detection across both modalities. Subsequently, L2-Net is trained to extract feature descriptors, producing score maps, orientation maps, and scale maps that encode the spatial positions, orientations, and scales of detected keypoints. The SuperGlue network then performs feature matching, filtering correct correspondences while rejecting outliers. Finally, affine transformation parameters are computed and bilinear interpolation is applied to achieve high-precision image alignment. Extensive experiments on a diverse remote-sensing dataset demonstrate that our method achieves superior matching performance over state-of-the-art approaches, with a correct match rate of 92.45% and root-mean-square error of 2.898. The source code and dataset will be made publicly available upon publication.