Multi-level cross-knowledge fusion with edge guidance for camouflaged object detection

Camouflaged object detection aims to identify objects that are “perfectly” assimilated into their surroundings, which has a wide range of valuable applications. The key challenge is that there exists high intrinsic similarities between the candidate objects and noise background. Despite the numerous learning frameworks developed in recent years to address this issue, they still struggle when confronted with highly deceptive camouflaged objects. In response, we propose a coarse-to-fine framework that leverages rough target positioning guided by edge semantics and refined identification assisted by cross-knowledge. Specifically, we dynamically mine additional object-related edge cues to guide representation learning in camouflaged object detection. This process focuses on exploring the initial object structure and facilitating the search for coarse regions of objects across multiple observation dimensions, mirroring biological mechanisms. Subsequently, multi-level cross-knowledge, refined through aggregation, aids in decoding the representation of precise regions. Employing a deep supervision strategy from top to bottom, our framework achieves accurate camouflaged object detection. Experiments conducted on various widely-used benchmark datasets demonstrate that our proposed network outperforms previous methods both qualitatively and quantitatively.