Abstract: In this paper, a registration method that integrates an edge segmentation network with feature point detection and description algorithms is presented for addressing the registration challenge posed by speckle noise and local distortions in multi-modal remote sensing images. Firstly, an improved feature extraction operator is employed to extract strong edge features of SAR images. Then, strong edge feature labels are constructed, an improved Deeplabv3+ edge segmentation model is trained to extract strong edge features from the images in a deep network manner. Finally, the proposed algorithm is used to detect and describe feature points on the feature maps. By integrating a deep learning-based semantic segmentation algorithm with traditional robust feature point detection and description methods, the reliability and the robustness of the algorithm have been improved. Registration tests involving translation, rotation, and scaling transformations are carried out on four types of images, and the results show that the aglorithm′s average RMSE reaches 2.088, demonstrating the superiority of the proposed algorithm.