FLRNet: A bio-inspired three-stage network for Camouflaged Object Detection via filtering, localization and refinement

Yilin Zhao, Qing Zhang, Yuetong Li

Abstract: Camouflaged Object Detection (COD) aims to segment camouflaged objects, which poses a more challenging task than generic object detection due to the high intrinsic similarity between the foreground and background. In this paper, we propose a bio-inspired three-stage network to uniformly highlight the complete camouflaged objects with explicit boundaries by information filtering, global localization, and progressive refinement. Specifically, we propose the local filtering module (LFM) to mimic the filtering stage by eliminating redundant information using cross-level information to reduce background interference. Furthermore, we design a global localization module (GLM) to determine positions of potential camouflaged objects by directly and progressively aggregating filtered features across different levels. Finally, to fully capture discriminative subtle features, we employ the progressive refinement module (PRM) to distinguish the details and structures by adopting a dual-attention strategy, progressively correcting prediction errors. Quantitative and qualitative experimental results demonstrate that our proposed FLRNet outperforms the state-of-the-art COD methods.