Main body, edge decomposition and reorganization network for building change detection

Traditional neural network methods for building change detection tend to produce the saw-tooth boundaries, and they are difficult to accurately identify change boundaries in dense building areas. To address that, this paper proposes a change detection method based on main body, edge decomposition and reorganization network. The proposed method performs change detection by respectively modeling body and edges features of buildings, which is on the basis on the characteristics of strong similarity between the body pixels and weak similarity between the edge pixels. In the definition of the proposed method, we first yield dual-temporal multi-scale difference features using a Siamese ResNet structure, and then separate the body features and edge features of buildings by learning a flow field. Subsequently, a feature optimization structure is designed to refine the body and edge features using the body and edge tags. Finally, the optimized body and edge features are reorganized to generate an end-to-end change detection model. Experiments have been performed by using the publicly available building dataset LEVIR-CD, and the results show that the proposed method can accurately identify the boundaries of changing buildings, and obtain better results compared with the methods based on U-Net network and these combining spatial-temporal attention.