Boundary-and-object collaborative learning network for camouflaged object detection

Existing camouflaged object detection (COD) approaches have achieved remarkable success in detecting and segmenting camouflaged objects that visually blend into the surroundings. However, there are still some challenging and critical issues, including inaccurate localization of target objects with varying scales, and incomplete identification of subtle details. To address these problems, we propose a novel boundary-and-object collaborative learning network (BCLNet) for camouflaged object detection, which simultaneously extracts and progressively refines the position and detail information to ensure segmentation results with uniform interiors and clear boundaries. Specifically, we design the Adaptive Feature Learning (AFL) module to generate the boundary information for identifying the details and the object information for positioning the target objects, and then optimize the two types of features in an interactive learning manner. In this way, the boundary feature and the object feature are able to learn from each other and compensate deficiencies for themselves, thus improving the semantic and detail representation. Moreover, to fully explore the complementarity between the cross-level features, we propose the Boundary-guided Selective Fusion (BSF) module to introduce the boundary cue to help the cross-level feature integration, enriching the semantic information while preserving the detail information. Extensive experimental results demonstrate that our BCLNet outperforms the state-ofthe-art COD methods on four widely used datasets. The link to our code and prediction maps are available at https://github.com/ZhangQing0329/BCLNet.