Finetuning Deep Neural Networks for SAR Image Registration

The scope of Synthetic Aperture Radar (SAR) image registration is rapidly expanding beyond traditional multi-modal applications to include emerging domains such as SAR odometry, navigation, and SAR-based SLAM, where accurate registration between sequential SAR images is essential. In this work, we explore the feasibility of using deep neural network (DNN) featurematching models for SAR-to-SAR image registration. A new dataset of SAR image pairs was constructed to facilitate training and evaluation. Three state-of-the-art DNN models-ROMA, SuperGlue, and ELoFTR-were tested. ROMA, a dense matcher, achieved high accuracy without additional training, demonstrating strong generalization. In contrast, SuperGlue and ELoFTR performed poorly with pretrained weights but showed substantial improvement after fine-tuning on the SAR dataset. SuperGlue's rotation RMSE decreased by 35.3% (from 0.3265° to 0.2111°), and x-translation error dropped by 55.5% (from 6.70 m to 2.9797 m). ELoFTR exhibited even greater gains, with an 82.1% reduction in rotation RMSE and over 95% improvement in xtranslation accuracy. All models achieved sub-meter accuracy with sub-second inference times, demonstrating the potential of fine-tuned DNN matchers for real-time SAR-SAR registration tasks.