Detecting camouflaged objects via cross-level context supplement

Camouflaged object detection (COD) is to distinguish the target objects with varied sizes and shapes from the low-contrast real-world scenarios. Although deep learning-based methods have made great progress, it is still challenging to accurately detect and segment the complete and edge-preserving camouflaged objects. In this paper, we propose a novel cross-level context supplement network (CCSNet) to effectively utilize cross-level features to provide additional information, which can compensate for the deficiencies of the current level potentials. Specifically, we develop a selective cross-level aggregation (SCA) module to fully explore the cross-level different but complementary cues to detect camouflaged objects with different scales. It makes each level of the network adaptively focus on the informative features with the assist of the channel dependencies and spatial relationship provided by the adjacent levels. Furthermore, considering that the camouflaged objects are hardly distinguishable from the backgrounds, we design a location and boundary supplement (LBS) module to directly incorporate the global and edge information to different levels, thus enhancing the spatial coherence of interior regions and reducing the uncertainty of boundary regions. Comprehensive experiments are conducted on four public datasets to demonstrate the effectiveness of our CCSNet. In addition, we apply our model to two other dense pixel prediction tasks to further demonstrate the effectiveness and generalization ability of our network. The code, trained model and predicted maps will be available upon acceptance of this paper for publication.