Remote Sensing Image Change Detection based on Swin Transformer and Cross-Attention Mechanism

Abstract With the advancement of Vision Transformer (ViT) in remote sensing image change detection, the most popular feature extraction methods involve using pre-trained ResNet or VGG networks. However, due to differences between the pre-trained datasets and remote sensing datasets, using ResNet or VGG as the backbone network does not yield optimal change detection results. Furthermore, current methods fall short in delivering comprehensive change detection regions and accurately delineating irregular change boundaries. To address these challenges, we propose using the Swin Transformer as our backbone network. This approach not only facilitates training on our custom remote sensing datasets but also enables the extraction of multi-scale and multi-level visual features. We also utilize the Cross-Attention mechanism to extract differential features and consider the interactive information between the two-branch features. By leveraging the attention mechanism, we can suppress non-changing information while effectively extracting changing information, thereby enhancing change detection results by prioritizing relevant changes. Additionally, we introduce a cascading feature fusion method to integrate features across different scales. This approach effectively mitigates information loss during fusion and maximizes the utilization of the extracted multi-scale features. Extensive experiments demonstrate that our proposed approach achieves state-of-the-art performance on three widely used change detection datasets.