Bidirectional Symmetry Network with Dual-Field Cyclic Attention for Multi-Temporal Aerial Remote Sensing Image Registration

Multi-temporal remote sensing image registration is a geometric symmetry process that involves matching a source image with a target image. To improve the accuracy and enhance the robustness of the algorithm, this study proposes an end-to-end registration network—a bidirectional symmetry network based on dual-field cyclic attention for multi-temporal remote sensing image registration, which mainly improves feature extraction and feature matching. (1) We propose a feature extraction framework combining an attention module and a pre-training model, which can accurately locate important areas in images and quickly extract features. Not only is the dual receptive field module designed to enhance attention in the spatial region, a loop structure is also used to improve the network model and improve overall accuracy. (2) Matching has not only directivity but also symmetry. We design a symmetric network of two-way matching to reduce the registration deviation caused by one-way matching and use a Pearson correlation method to improve the cross-correlation matching and enhance the robustness of the matching relation. In contrast with two traditional methods and three deep learning-based algorithms, the proposed approach works well under five indicators in three public multi-temporal datasets. Notably, in the case of the Aerial Image Dataset, the accuracy of the proposed method is improved by 39.8% compared with the Two-stream Ensemble method under a PCK (Percentage of Correct Keypoints) index of 0.05. When the PCK index is 0.03, accuracy increases by 46.8%, and increases by 18.7% under a PCK index of 0.01. Additionally, when adding the innovation points in feature extraction into the basic network CNNGeo (Convolutional Neural Network Architecture for Geometric Matching), accuracy is increased by 36.7% under 0.05 PCK, 18.2% under 0.03 PCK, and 8.4% under 0.01 PCK. Meanwhile, by adding the innovation points in feature matching into CNNGeo, accuracy is improved by 16.4% under 0.05 PCK, 9.1% under 0.03 PCK, and 5.2% under 0.01 PCK. In most cases, this paper reports high registration accuracy and efficiency for multi-temporal remote sensing image registration.