DRRNet: Macro-Micro Feature Fusion and Dual Reverse Refinement for Camouflaged Object Detection

TL;DR

DRRNet tackles camouflaged object detection by jointly modeling panoramic context and fine-grained local details through a four-stage pipeline: panoramic perception, detail mining, cross-level fusion, and dual reverse calibration. The architecture combines an Omni-Context Module (OCM) and Micro-Detail Module (MDM) in parallel branches, fused via Macro-Micro Fusion (MMF), with a Global Rough Decoder (GRD) producing a coarse map and a Dual Reverse Refinement Module (DRRM) performing two-stage spatial-frequency refinement. Key contributions include the OCM/MDM for multi-scale context and local detail capture, MMF for cross-domain fusion, and DRRM for progressive refinement, all validated on COD benchmarks and extended to polyp segmentation. The results demonstrate state-of-the-art COD performance with efficient compute, and the approach shows cross-task adaptability to medical image segmentation tasks, highlighting practical impact for edge devices and complex scene understanding.

Abstract

The core challenge in Camouflage Object Detection (COD) lies in the indistinguishable similarity between targets and backgrounds in terms of color, texture, and shape. This causes existing methods to either lose edge details (such as hair-like fine structures) due to over-reliance on global semantic information or be disturbed by similar backgrounds (such as vegetation patterns) when relying solely on local features. We propose DRRNet, a four-stage architecture characterized by a "context-detail-fusion-refinement" pipeline to address these issues. Specifically, we introduce an Omni-Context Feature Extraction Module to capture global camouflage patterns and a Local Detail Extraction Module to supplement microstructural information for the full-scene context module. We then design a module for forming dual representations of scene understanding and structural awareness, which fuses panoramic features and local features across various scales. In the decoder, we also introduce a reverse refinement module that leverages spatial edge priors and frequency-domain noise suppression to perform a two-stage inverse refinement of the output. By applying two successive rounds of inverse refinement, the model effectively suppresses background interference and enhances the continuity of object boundaries. Experimental results demonstrate that DRRNet significantly outperforms state-of-the-art methods on benchmark datasets. Our code is available at https://github.com/jerrySunning/DRRNet.

Figures (10)

• Figure 1: Visual comparison of models utilizing various types of information. (b) Baseline model. (c) Convolution-based model, which fails to accurately detect regions where the crab's legs closely blend with the environment. (d) Transformer-based model, which yields acceptable results, yet remains deficient in local detail. (e) Our model, employing a four-stage framework of panoramic perception – detail mining – cross-level fusion – dual reverse calibration, produces the best results.
• Figure 2: The framework of the proposed DRRNet.OCM extracts panoramic contextual semantics, MDM extracts local detail features, and MMF is used to fuse global and local features. GRD is responsible for generating a coarse prediction map, which serves as the prior for precise predictions. DRRM utilizes spatial edge prior information and frequency-domain noise suppression for dual reverse refinement of the results.
• Figure 3: Details of the proposed OmniContext Module (left) and MicroDetail Module (right).
• Figure 4: The details of Macro-Micro Fusion(MMF) module.
• Figure 5: The details of Global Rough Decoder(GRD) module.