PAGNet: a perception-aware attention-guided network for building change detection in remote sensing images

In view of the challenges of strong background interference, weak feature fusion and insufficient multi-scale semantic perception in the task of building change detection in remote sensing images, this paper proposes a change detection method based on the Perception-Aware Attention-Guided Network (PAGNet). The network adopts a typical encoder-decoder structure and integrates three key designs: the first is the BAM-guided Perception Enhancement Module (BPAEM), which introduces contextual perception and channel attention information by constructing a change guidance path, effectively strengthening the attention to the change target and suppressing pseudo-changes; the second is the Mixed Attention Module (MAM), which combines the spatial and contextual interaction mechanisms to improve the modeling ability of fine-grained change areas in the decoding stage; the third is the design of a composite loss function that integrates adaptive focal loss and boundary-aware dice loss to alleviate category imbalance and improve boundary detection accuracy, thereby enhancing overall discrimination ability. By realizing deep fusion of cross-temporal features in a multiscale semantic space, PAGNet can accurately distinguish real change areas in complex backgrounds. Experimental results show that the proposed method achieves excellent performance on public change detection datasets.