AFSNet: Adaptive Feature Suppression Network for Remote Sensing Image Change Detection

Recently, deep learning has made great progress in remote sensing change detection. However, some interference information are involved in bitemporal images which causes the algorithms to be affected by pseudo-changes in the background, such as shooting angle, seasonal turnover, and illumination intensity. Some researchers try to use the attention mechanism to solve this issue. To date, the existing attention methods explore incompletely the potentiality of feature suppression. Unlike existing spatial attention methods, we hope to obtain the interested features while removing some irrelevant-task features. From this perspective, we propose a new change detection architecture, i.e. adaptive feature suppression network (AFSNet), which includes two core components: adaptive feature suppression attention (AFSA) module and spatial and channel feature fusion (SCFF) strategy. We carefully design the AFSA inspired by soft threshold function, and it only uses 10 parameters to suppress interference information. Specifically, we remove spatial irrelevant information in the calculated process of soft threshold function and introduce a set of scaling factors to restrain redundant channel features. SCFF is an effective feature fusion strategy, and it utilizes simultaneously learnable addition and concatenation operations to aggregate better bitemporal features. Compared with some state-of-the-art (SOTA) methods on two challenging remote sensing change detection datasets, ASFNet can achieve superior performance. The code will be publicly available at https://github.com/tlyslll/AFSNet.