Abstract: In co-salient detection, due to the neglect of the importance of global attention in extracting common features and the neglect of the use of multi-scale features, the existing algorithms still have some problems, such as lack of semantic features and poor prediction in multi-significant target scenarios. To solve the above problems, this paper proposes a new co-salient detection model. The model is mainly composed of a global consensus generation module and a consensus refinement module. The global consensus generation module is designed to globally capture the common characteristic, specifically consistency, across the entire set of images. Meanwhile, the consensus refinement module is primarily utilized to refine the initially extracted consistency, effectively averting misdetection in the presence of multiple targets. The test results on three classical open data sets show that the network is better than the 13 most advanced methods in M_a, F_\mathrmm,E_\mathrmm,S_\mathrmm , and it also demonstrates better performance of collaborative saliency information detection. Additionally, the efficacy of the newly introduced module is validated through ablation experiments.