Bidirectional Class-Text and Vision Interaction Network for Camouflaged Object Detection

Camouflaged object detection (COD) aims to identify targets that are highly integrated into their surroundings. Existing approaches often heavily rely on immediate visual cues, leading to inadequate feature representations, especially in complex scenarios. Recent works attempt to incorporate textual information have shown potential. However, they typically depend on large-model-generated descriptions or treat vision-language models as static encoders, deferring cross-modal alignment to the decoding stage. These approaches are prone to inaccuracy and semantic bias. To overcome these limitations, we propose a novel COD framework that establishes bidirectional interactions between class-level text and visual features. By leveraging DINOv2’s visual perception and CLIP’s textual semantics, our method facilitates progressive cross-modal alignment from global to pixel levels, significantly enhancing detection stability and accuracy in challenging camouflage environments. Extensive experiments on three public datasets demonstrate the effectiveness and superiority of our proposed network.