Source-linked AI summary
Frequency Perception Network for Camouflaged Object Detection
Runmin Cong, Mengyao Sun, Sanyi Zhang, Xiaofei Zhou, Wei Zhang, Yao Zhao
TL;DR
COD remains difficult because camouflaged objects resemble their surroundings, while existing methods mainly localize them in RGB space. FPNet learns frequency-aware features for coarse localization and progressively refines predictions using prior correction, cross-layer fusion, and high-resolution boundary information. It achieves competitive qualitative and quantitative performance on three benchmark datasets.
Problem
COD is challenging because camouflaged objects have low contrast with their backgrounds, and existing methods mainly operate in RGB space in difficult scenarios.
Method
FPNet combines a learnable frequency-perception module for coarse localization with a detail-preserving refinement mechanism using prior correction, cross-layer association, and high-resolution features.
Results
FPNet achieves competitive qualitative and quantitative performance on the COD10K, CHAMELEON, and CAMO benchmark datasets.
Takeaways & Limitations
The results support combining RGB-aware and frequency-aware clues in a unified two-stage solution for challenging camouflaged object detection.
Takeaways & Limitations
The coarse localization stage does not provide sufficient integrity and accuracy, requiring a detail-preserving fine localization mechanism.
Abstract
from arXiv · showhide
Camouflaged object detection (COD) aims to accurately detect objects hidden in the surrounding environment. However, the existing COD methods mainly locate camouflaged objects in the RGB domain, their performance has not been fully exploited in many challenging scenarios. Considering that the features of the camouflaged object and the background are more discriminative in the frequency domain, we propose a novel learnable and separable frequency perception mechanism driven by the semantic hierarchy in the frequency domain. Our entire network adopts a two-stage model, including a frequency-guided coarse localization stage and a detail-preserving fine localization stage. With the multi-level features extracted by the backbone, we design a flexible frequency perception module based on octave convolution for coarse positioning. Then, we design the correction fusion module to step-by-step integrate the high-level features through the prior-guided correction and cross-layer feature channel association, and finally combine them with the shallow features to achieve the detailed correction of the camouflaged objects. Compared with the currently existing models, our proposed method achieves competitive performance in three popular benchmark datasets both qualitatively and quantitatively.
1 INTRODUCTION
COD is difficult because camouflaged objects resemble their backgrounds and vary in appearance, while RGB-domain methods remain challenged by boundaries, multiplicity, and occlusion. FPNet addresses these issues with frequency-aware coarse localization followed by progressive detail refinement.
- COD is challenging because objects have low contrast with backgrounds and may vary in shape, size, and texture.
- Existing COD methods struggle with multiple objects, uncertain boundaries, and occlusion while mainly seeking discriminative texture inconsistencies in RGB images.
- FPNet is a two-stage framework that combines RGB and frequency domains for camouflaged object detection and achieves competitive performance on COD10K, CHAMELEON, and CAMO.
- The frequency-perception module automatically learns high- and low-frequency features to distinguish camouflaged objects from backgrounds and produce a coarse localization.
- A progressive refinement mechanism uses prior-guided correction, cross-layer channel association, and shallow high-resolution features to refine object boundaries.
2 RELATED WORK
COD research has progressed from handcrafted features and adapted SOD models to customized CNN and Transformer networks, including two-stage, multitask, edge-guided, and frequency-guided strategies. FPNet extends this direction with a two-stage architecture that uses learned frequency features for initial positioning.
- Earlier COD methods used handcrafted color, statistical, intensity, optical-flow, and texture features, while simple SOD retraining was insufficient for COD.
- Recent CNN-based COD methods commonly use two-stage processing, multitask learning, or additional cues such as frequency information.
- Frequency-domain studies have used DCT coefficients, spatial-to-frequency network conversion, and modeled frequency interactions, but offline frequency extraction can limit flexibility.
- FPNet contributes a feature-extraction backbone, a frequency-guided coarse localization stage, and a detail-preserving fine localization stage.
- The proposed frequency-domain positioning uses high- and low-frequency information to find breakthrough points for locating camouflaged objects.
3 OUR APPROACH
FPNet combines RGB and frequency-domain features in a two-stage pipeline: frequency-guided coarse localization followed by detail-preserving fine localization. Its modules learn frequency components, progressively correct high-level features, and restore boundary detail with shallow high-resolution features.
- Overview: FPNet uses a frequency-guided coarse stage and a detail-preserving fine stage to exploit complementary RGB and frequency-domain information for camouflaged object detection.The network extracts pyramid features, localizes objects coarsely from frequency-aware representations, then refines predictions through correction and boundary-detail processing.
- Frequency-guided Coarse Positioning: Octave convolution automatically separates features into high- and low-frequency components, representing rapidly changing textures and gentler intensity structures for frequency-domain object representation.The module learns the decomposition online rather than relying on offline DCT extraction; the two components are fused because texture and contour attributes both matter for localization.
- Frequency-guided Coarse Positioning: Neighbor-layer decoding progressively integrates the top-three frequency-domain features and produces a coarse mask that reveals the initial camouflaged-object location.The decoder uses cross-layer semantic context through neighbor connections before a convolution generates the coarse prediction S1.
- Detail-preserving Fine Localization: The correction fusion module refines adjacent-layer features by combining a coarse mask with prior-guided correction and channel-level association modeling.The coarse mask selects object-related features, while channel interactions transfer high-level information across layers and residual processing generates cross-level fusion features.
- Detail-preserving Fine Localization: RFB and spatial attention enhance first-layer high-resolution features before fusion with CFM output, refining camouflaged-object boundaries into the final prediction.The training objective combines weighted binary cross-entropy and weighted IoU losses for the coarse, intermediate, and final maps.
4 EXPERIMENT
Experiments evaluate FPNet on three benchmark datasets against state-of-the-art methods and through ablations of its frequency-aware, high-resolution, and correction-fusion components.
- Quantitative Evaluation: FPNet outperforms all compared state-of-the-art models across COD10K, CAMO, and CHAMELEON evaluation metrics.The comparison includes 12 state-of-the-art methods and reports quantitative results in Table 1.
- Quantitative Evaluation: The weighted F-measure gains over the second-best method are 2.6% on COD10K-Test, 7.2% on CAMO-Test, and 1.3% on CHAMELEON.On CAMO-Test, FPNet also reaches Fωβ=0.806 and improves MAE over ZoomNet by 15.2%.
- Effectiveness of Frequency Perception Module: Adding the frequency perception module improves every metric over the PVT-only baseline by learning high- and low-frequency features online for coarse localization.The module uses octave convolution and is described as flexible enough to integrate into other frameworks.
- Effectiveness of High-resolution Preserving Module: High-resolution preserving improves PVT+FPM by refining boundaries that high-level semantic features inadequately capture across varying object scales.The fine-localization stage uses low-level RGB-aware high-resolution features to recover detailed boundaries.
- Effectiveness of Correction Fusion Module: The correction fusion module further improves all metric scores by combining the coarse mask with neighboring-layer channel correlations.The coarse prediction supplies object-centric guidance while channel-wise correlation integrates adjacent features.
- Detail Analysis of the Frequency-aware Information: High-frequency features outperform low-frequency features alone, while their fusion produces more accurate masks by combining key-object cues with more complete regions.Qualitative analysis shows high-frequency predictions capture key object parts, whereas low-frequency predictions include more background interference.
5 CONCLUSION
The paper concludes that FPNet combines RGB and frequency domains through a two-stage localization framework for camouflaged object detection. Comparisons and ablations on three benchmark datasets validate the approach, while long-tail recognition remains a future direction.
- 5 CONCLUSION: FPNet automatically separates frequency information for coarse-mask prediction, then uses detail-preserving refinement and correction fusion to improve the coarse map.The framework integrates RGB-aware and frequency-aware processing in two stages.
- 5 CONCLUSION: Comprehensive comparisons and ablation studies on three benchmark COD datasets validate the effectiveness of FPNet.
- 5 CONCLUSION: The paper identifies long-tail recognition as a direction for future COD research.
A.1 Qualitative Comparison with SOTA Methods
Qualitative comparisons show FPNet remains competitive in multiple-object, fine-object, occluded, and complex-background scenarios. Its predictions preserve object boundaries and small details while suppressing distracting background structures.
- A.1 Qualitative Comparison with SOTA Methods: FPNet predicts all camouflaged objects with clear, sharp boundaries when scenes contain multiple objects.Other methods may miss objects or produce unclear boundaries in these cases.
- A.1 Qualitative Comparison with SOTA Methods: FPNet captures fine burr-like structures and other trivial details that competing methods fail to represent fully.The comparison also reports advantages when camouflaged targets are occluded.
- A.1 Qualitative Comparison with SOTA Methods: In complex backgrounds, FPNet detects camouflaged targets while rejecting misleading shadows and other interference.The ghost pipefish example contrasts FPNet with methods that mistake indistinguishable shadows for camouflaged objects.
A.2 Visualization of Ablation Studies
The ablation visualizations show progressively improved localization, structure, and boundaries as frequency perception, high-resolution preservation, and correction fusion are added.
- A.2 Visualization of Ablation Studies: Correction fusion achieves the best qualitative result, with accurate localization, complete structure, and sharp boundaries.The baseline roughly identifies the object, while the final design resolves remaining structural and boundary errors.
- A.2 Visualization of Ablation Studies: Frequency perception improves the camouflaged human’s shoulder region, but leg integrity remains unresolved.
- A.2 Visualization of Ablation Studies: High-resolution preservation completes more leg details but introduces noise before correction fusion produces the final refined result.
- A.2 Visualization of Ablation Studies: Figure 9 compares the input, ground truth, full model, and ablation variants for qualitative assessment.