Transformer-based network with difference information aware attention module

Abstract

Existing great progress has been made through remote sensing image change detection (CD) methods, which use convolutions and transformations appropriately. However, due to the complexity of the objects photographed on satellites, the out-of-phase of the same object may exhibit a fine feature diversity and a wide range of variations. Meanwhile, in the previous work, pure convolutional networks and Transformer networks that focus on local information perform better on non-local information. Therefore, current methods still have the problem of insufficient discrimination between irregular changes and diverse boundaries. Here, to solve this biphasic dilemma, we put forward a network that uses the difference information between bitemporal images as a guide to refine features. We use one feature map to subtract another and use multi-scale convolution to refine the difference map. We then designed an attention module similar to self-attention using differential information. In addition, considering the need for both local and global information in change detection, we use the full Swin Transformer framework with shift window self-attention to pass information between local areas. In addition, in order to enhance the network’s ability to recognize, members of our team reformulate supervised, traditional saliency map prediction through saliency and edge prediction. Notably, our proposed model significantly outperforms existing single-branch methods and state-of-the-art performance was achieved on four famous public datasets in terms of the F1-score index, interest in joint indices, and population indices.