DLCUNet: A Densely Local Calibrate U-shape Net for Change Detection in Optical Remote Sensing Images

The task of change detection in remote sensing images aims to capture difference information from time-phase sequence images and provide references for human activities. Because change detection tasks have similar goals to image segmentation tasks, researchers often use semantic segmentation techniques to implement change detection tasks. This approach has greatly improved the robustness and efficiency of change detection compared to the past, but there are still three difficulties: the segmentation map is not continuous, the multi-scale target is not salient, and the attention mechanism is too complicated. In order to deal with these three difficulties, this paper designs a new encoder unit using the idea of local information embedding, improves the decoder unit with a non-training feature fusion method, and combines the topological structure of UNet++ to propose a new change Detect network. The new network has strong anti-noise ability and multi-scale target acquisition ability, and is more concise. The method in this paper has been verified on the CDD dataset and LEVIR-CD dataset.