SRTPN: Scale and Rotation Transform Prediction Net for Multimodal Remote Sensing Image Registration

How to recover geometric transformations is one of the most challenging issues in image registration. To alleviate effect large distortion multimodal remote sensing registration, a scale and rotate transform prediction net proposed this paper. First, reduce between reference sensed images, regression module constructed via CNN feature extraction FFT correlation, can be recovered roughly. Second, rotation estimate developed for predicting angles scale-recovered images. Finally, obtain accurate registration results, LoFTR employed match geometric-recovered The network was evaluated on GoogleEarth, HRMS, VIS-NIR UAV datasets with contrast differences distortions. experimental results show that number correct matches our model reached 74.6%, RMSE achieved 1.236, which superior related methods.