RegCDNet: A RegNet-Based Framework for Remote Sensing Image Change Detection Combining Feature Enhancement and Gating Mechanism

With the rapid development of remote sensing (RS) technology, it has become more and more convenient to obtain multi-temporal RS images, which provides new opportunities for the research and development of change detection (CD) technology. However, existing methods still have shortcomings in recognizing complex targets and accurately distinguishing the edges of change regions when performing CD in complex backgrounds. Therefore, we propose a novel CD model called RegCDNet, which is specifically designed to address the needs of RS image CD. The model employs RegNet as the backbone network for feature extraction, using a simple and efficient strategy to fuse shallow features. We designed an ac-dramit feature enhancement module (AFEM) that combines atrous convolution and a transformer containing a dual attention mechanism, which can efficiently capture long-range dependencies between pixels while focusing on local information. A dual-gated fusion module (DGFM) is also designed, which employs a dual gating mechanism for feature fusion and can dynamically adjust the fusion weights. The experiment achieved 91.22%, 92.84% and 82.52% F1 metrics on the LEVIR-CD, WHU-CD and SYSU-CD datasets, respectively.