Transformer Self-Attention Change Detection Network with Frozen Parameters
Peiyang Cheng, Min Xia, Dehao Wang, Haifeng Lin, Zikai Zhao
Abstract
The purpose of change detection is to recognize changed areas from a pair of two remote sensing images. However, since change areas often include multiple terrain features, this demands enhanced feature extraction capability from the model. This paper proposes a frozen-parameter Transformer self-attention change detection network (ZAQNet). The network integrates four innovative modules: a GIAU (Generalized Image Attention Unit) which can effectively fuse the features of two remote sensing images and accurately focus on changing areas; a GSAU (Global Spatial Attention Unit) which performs self attention processing in the image spatial dimension to enhance the model’s ability to capture global change information; a GSCU (Global Semantic Context Unit) which performs self-attention operations in the channel dimension to enhance the model’s attention to feature maps containing changing information; and a PRU (Patch Refinement Unit) which extracts and refines spatial position information from the underlying feature map, optimizing the restoration effect of the feature map. The experiments on the BTRS-CD and LEVIR-CD datasets show that ZAQNet performs excellently in change detection tasks. Among them, the change detection index F1 and IOU are better than the comparison model. These results fully demonstrate the superiority, robustness, and generalization ability of ZAQNet in change detection tasks and provide an efficient and reliable solution for remote sensing image analysis.