Multi-scale Refocusing Attention Siamese Network

Deep learning has achieved significant success in change detection due to its ability to automatically extract complex features. Recent research has focused on utilizing attention mechanisms. However, most attention mechanisms still struggle to fully exploit the local and global contextual relationships and often suffer from high computational complexity and lack robustness against pseudo-changes. Therefore, this paper proposes a method called Multi-scale Refocused Attention Siamese network, which captures change regions through multi-scale attention mechanisms and enhances model with prior knowledge for complex environments, thereby improving change detection capability. Experimental results demonstrate that the proposed method achieves F1 scores of 95.9% and 90.3% on two commonly used change detection datasets, CDD and WHU-CD respectively, proving its effectiveness and superiority in enhancing change detection performance.