Co-Saliency Detection Based on Feature Enhancement and Contrast Learning

Co-saliency object detection aims to identify commonly occurring salient objects from a set of related images. The challenges lie in (1) reducing the influence of non-collaborative salient objects that are only salient in individual images, not in the group of images; and (2) suppressing the influence of background noise in complex scenes. A co-saliency detection method based on feature enhancement and foreground background comparison learning is proposed to address the above difficulties. Firstly, the shared attributes of the images in the group are obtained by calculating the correlation of the images, resulting in co-saliency prototypes that contain the position and semantic information of the co-salient objects. Then, the co-saliency prototypes are used to enhance the co-salient objects in the deep features from the encoder, while suppressing the influence of non-co-salient objects. In addition, foreground-background contrast learning is used during the training phase to enhance the similarity between foreground objects and co-saliency features, as well as the difference between background regions and co-saliency features, in order to emphasize the features of foreground objects and reduce the influence of background regions. Extensive experiments are conducted on three publicly available datasets CoCA, CoSOD3k, and Cosal2015 to compare with existing state-of-the-art algorithms. The results indicate that the proposed algorithm can effectively improve the detection accuracy of co-salient objects in complex scenes while reducing the false detection rate caused by non-co-salient salient objects.