Camouflaged object detection via dual domain fusion

In camouflaged object detection (COD), the goal of accurately and completely segmenting foreground objects from the background on a per-pixel basis is pursued by various network architectures. Despite their specialized module designs tailored for camouflaged object detection, most of these approaches operate primarily within the RGB domain to locate camouflaged objects. Given the intrinsic characteristics of camouflaged objects, solely segmenting them within the RGB domain often encounters significant interference and challenges. To address this issue and leverage the advantages of detecting camouflaged objects from the frequency domain, this paper introduces a Dual-Domain Fusion Network (DDFNet). By integrating frequency domain features with RGB domain features, DDFNet exploits the complementary strengths of both domains, achieving precise localization and segmentation of camouflaged objects. DDFNet is composed of three main modules. The MHFM performs group fusion of backbone features to enhance the representation of these features. The GFGM, based on a dual-phase architecture, extracts the camouflaged target hidden in the RGB domain by operating in the frequency domain. It achieves this by leveraging the detailed information contained in high-frequency features and the spatial information in low-frequency features, thereby obtaining a frequency-domain feature representation of the camouflaged target, which supplements the RGB domain features. Finally, through GHIM, DDFNet performs complementary fusion of the frequency-domain features and the RGB-domain features, utilizing their respective advantages to achieve precise localization of the camouflaged target. Compared to other state-of-the-art methods in camouflaged object detection, DDFNet demonstrates superior performance. Additionally, we introduce polyp segmentation as a downstream task for the proposed network to showcase the ability of DDFNet in solving real-world problems.