Abstract : In the case of limited samples and complex scenes (seasonal changes, brightness changes, building renovations, etc.), there are still many limitations in remote sensing change detection. Directly explore the direct relationship between dual-time images before doing difference features on the two images. The feasibility of interaction. In this paper, we propose a dual-branch multi-level spanning temporal network (DMINet) to efficiently derive change detection. Specifically, by unifying self-attention (SelfAtt) and cross-attention (CrossAtt) in one module, we propose an intertemporal joint attention (JointAtt) block to guide the global features between them, motivating intra-layer representations between information coupling while suppressing task-irrelevant interference. In addition, centering on the detection of differential features, based on a simple backbone network without complex structures, such as ResNet, two issues are focused on, differential acquisition using subtraction and concatenation and multi-level differential aggregation using incremental feature rearrangement.
Table of contents