Registration of Multiresolution Remote Sensing Images Based on L2-Siamese Model

The registration of multiresolution optical remote sensing images has been widely used in image fusion, change detection, and image stitching. However, traditional registration methods achieve poor accuracy in the registration of multiresolution remote sensing images. In this study, we propose a framework for generating deep features via a deep residual encoder (DRE) fused with shallow features for multiresolution remote sensing image registration. Through an L2 normalization Siamese network (L2-Siamese) based on the DRE, the multiscale loss function is used to learn the attribute characteristics and distance characteristics of two key points and obtain the trained feature extractor. Finally, the DRE is used to extract the deep features of the key points and their neighbors, which are concatenated with the shallow features into a fusion feature vector to complete the image registration. We performed comprehensive experiments on four sets of multiresolution optical remote sensing images and two sets of synthetic aperture radar images. The results demonstrate that the proposed registration model can achieve subpixel registration. The relative registration accuracy improved by 1.6%-7.5%, whereas the overall performance improved by 4.5%-14.1%.