Abstract
Change detection is widely utilized across various domains, such as disaster monitoring, where it aids in identifying differences between images captured at different time intervals. However, current methods often lack constraints on intermediate features and fail to comprehensively model the temporal relationships among these features. Additionally, they rely on simplistic fusion mechanisms, leading to suboptimal network performance. In this paper, we propose: (1) a Feature Information Interaction Module (FIIM) based on spatial attention to enhance semantic information; (2) a Feature Pair Fusion Module (FPFM) with dual-branch structure to model bi-temporal relationships; and (3) a Multi-Scale Supervision Method (MSSM) using contrastive learning to better constrain intermediate features. Comparative experiments conducted on the CDD and LEVIR-CD datasets demonstrate the superiority of our proposed network over existing state-of-the-art methods. Code repository: https://github.com/joyeuxni/SNIIF-Net .