Dual-Branch Enhanced Network for Change Detection

Change detection is an essential task in intelligent monitoring, and the accuracy of detection is of central importance for subsequent target tracking and recognition. However, a series of challenges such as illumination change, severe weather, shadow, and camera jitter have brought great troubles. To reduce the impact of these factors, we propose a novel model, called dual-branch enhanced network (DBEN), which can simultaneously extract enough spatial features and context information. Specifically, we design a recurrent gated bottleneck module to get high-level features, and build the global attention module as an auxiliary branch to obtain fine resolution details. Moreover, we also propose a gated residual dense module to enhance feature expression by reconstructing the combined information. Meanwhile, a weighted loss function is designed to optimize the network. The proposed DBEN is verified on CDnet2014, DAVIS and AICD, which are three large-scale change detection datasets. Experimental results show that the proposed model is competitive in overall performance.