CCINet: A cascaded consensus interaction network for co-saliency object detection

Abstract

Co-saliency object detection imitates human attention behavior, with the aim of identifying common salient objects in a set of related images. Previous approaches generally suffer from a lack of interaction among the extracted co-saliency information. As a result, the detection maps often turn out to be incomplete or redundant. In this paper, we propose a Cascaded Consensus Interaction Network (CCINet) for co-saliency object detection. This network improves the fusion and interaction among features, thus making full use of the co-saliency information. In the encoding stage, we introduce an Edge Semantic Consensus (ESC) module. It effectively integrates low-level and high-level encoding information. In this way, it is able to capture both fine edge details and rich semantics. Meanwhile, the ESC module refines the co-saliency features, which enhances the detection of co-saliency regions. During the up-sampling stage, the Cascaded Contextual Aggregation (CCA) module employs attention mechanisms, adaptive pooling, and separated-dilated convolution for comprehensive feature extraction. This approach effectively reduces background noise and controls the number of parameters. Extensive experiments indicate that our model outperforms many excellent CoSOD methods in recent years on the three most popular benchmark datasets. Source code is available at: https://github.com/JoeLAL24/CCINet.git.