CJCNet: a cluster joint comparison network for co-saliency object detection
Longsheng Wei, Zhanxiang Zhou, Jiu Huang, Fan Xu · Journal of Electronic Imaging · 2025
Co-saliency object detection (CoSOD) mimics human attention mechanisms and is aimed at identifying common salient objects within a set of related images. Most of the previous methods usually overlooked taking into account both the intra-cluster consistency of the same type and the inter-cluster differences of different types. We propose a cluster joint comparison network for CoSOD, which conducts the search for co-saliency within the group and also maintains the differences among different types of objects. First, the intra-cluster search module locates common salient regions across multiple images in the same group to enhance intra-cluster consistency of co-saliency. Subsequently, the inter-cluster comparison module is used to increase the differences among salient objects of different categories, reducing the possibility that the model classifies objects with similar appearances but different types into one category. Next, the consistency discrimination classification module is used to map the input image to the output of the target type to assist in the classification of the detection results. Finally, the background of the predicted image is de-interfered through the context information extraction module. Extensive experiments show that our model achieves better performance than many existing CoSOD methods in recent years under the three most popular benchmark datasets.