Abstract: Co-saliency detection aims to identify the common and salient objects from a group of relevant images. The main challenge for co-saliency detection is how to mine and exploit the saliency cues of both intra-image and inter-image. A novel unified hierarchical neural network is presented. Firstly, the images are segmented by the superpixel segmentation algorithm, and the intra-image hierarchical saliency features are extracted to construct a graph model. Secondly, hierarchical salient graph embedding of the inter-image is mined to form a unified two-dimensional hierarchical feature system. Finally, a geometric attention module is further proposed in order to make full use of the intra-image and inter-image cues. The ablation experiments on the iCoSeg dataset show that each module in the proposed unified hierarchical neural network is effective. The maximum F-measure, mean absolute error and S-measure obtained with the proposed method on the iCoSeg dataset are 0.848 6, 0.107 6 and 0.813 4, respectively, which are comparable to or better than those with other 9 control methods. The highlight consistency and edges of the final obtained saliency map are significantly improved.