Source-linked AI summary
Context-aware Cross-level Fusion Network for Camouflaged Object Detection
Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, Nian Liu
TL;DR
Camouflaged object detection is challenged by low boundary contrast and substantial variation in object appearance. C2F-Net addresses this with attention-induced cross-level fusion followed by dual-branch global context modeling, and experiments report superior performance against state-of-the-art models on three benchmark datasets.
Problem
COD must identify and segment objects with extremely low boundary contrast and varied size and shape, while existing methods degrade on multiple-object and occluded scenes.
Method
C2F-Net cascades ACFM for attention-guided high-level cross-level feature fusion with DGCM for multi-scale global context modeling.
Results
C2F-Net outperforms 14 state-of-the-art models on three benchmark datasets across four evaluation metrics.
Takeaways & Limitations
The reported experiments support C2F-Net as an effective COD model for challenging camouflage scenarios.
Abstract
from arXiv · showhide
Camouflaged object detection (COD) is a challenging task due to the low boundary contrast between the object and its surroundings. In addition, the appearance of camouflaged objects varies significantly, e.g., object size and shape, aggravating the difficulties of accurate COD. In this paper, we propose a novel Context-aware Cross-level Fusion Network (C2F-Net) to address the challenging COD task. Specifically, we propose an Attention-induced Cross-level Fusion Module (ACFM) to integrate the multi-level features with informative attention coefficients. The fused features are then fed to the proposed Dual-branch Global Context Module (DGCM), which yields multi-scale feature representations for exploiting rich global context information. In C2F-Net, the two modules are conducted on high-level features using a cascaded manner. Extensive experiments on three widely used benchmark datasets demonstrate that our C2F-Net is an effective COD model and outperforms state-of-the-art models remarkably. Our code is publicly available at: https://github.com/thograce/C2FNet.
1 Introduction
Camouflaged object detection is difficult because objects have low boundary contrast and varied appearances, especially in scenes with multiple objects or occlusion. C2F-Net addresses these challenges by combining attention-guided cross-level fusion with global context modeling.
- Challenge: COD identifies and segments camouflaged objects whose boundaries have extremely low contrast with their surroundings.This makes objects difficult to discover and segment accurately.
- Challenge: Varied object size and shape further aggravate the difficulty of accurate COD.The paper highlights appearance variation as an additional challenge beyond boundary ambiguity.
- Research gap: Existing methods degrade on challenging cases such as multiple objects and occlusion, motivating richer context and effective cross-level feature fusion.The paper identifies both factors as important but often overlooked together by prior work.
- Proposed solution: C2F-Net fuses multi-level features with ACFM and then exploits global context using DGCM in a cascaded high-level-feature design.ACFM uses attention coefficients, while DGCM produces multi-scale representations for context exploitation.
- Reported contribution: C2F-Net outperforms 14 state-of-the-art models on three benchmark datasets across four evaluation metrics.This is the paper’s reported broad experimental conclusion.
2 Related Work
Related work covers camouflaged object detection and context-aware deep learning. Prior COD methods evolved from hand-crafted visual features toward deep learning, while context modeling uses multi-scale or dilated representations.
- Camouflaged Object Detection: Early COD methods relied on visual features such as color, texture, motion, and gradient, but hand-crafted features suited mainly simple scenarios.Integrating multiple features improved performance, yet limitations remained for real-world applications.
- Camouflaged Object Detection: Deep learning techniques were adopted for COD to address limitations of hand-crafted approaches.The supplied passage introduces this transition without detailing a specific deep-learning model.
- Context-aware Deep Learning: Contextual information supports object segmentation by enhancing feature representations for improved performance.Related approaches obtain context through multi-scale representations or dilated convolutions.
- Connection to C2F-Net: The proposed model’s architecture combines an attention-induced cross-level fusion module with a dual-branch global context module.Figure 2 presents these as the two key components of the overall architecture.
3 Proposed Method
C2F-Net extracts multi-level features, selectively fuses high-level features with attention, and enhances them through dual-branch global context modeling. Its training combines weighted BCE and weighted IoU losses to impose pixel-level and global-structure constraints.
- Overall architecture: C2F-Net extracts features from five Res2Net-50 layers and applies receptive field blocks to capture richer layer-specific features.The architecture uses five backbone feature levels and multi-branch receptive field blocks.
- Attention-induced Cross-level Fusion Module: ACFM uses Multi-Scale Channel Attention to compute informative coefficients for refining and fusing cross-level features.The module targets scale variation by integrating features according to attention cues.
- Attention-induced Cross-level Fusion Module: ACFM is applied only to high-level features f_i for i = 3, 4, 5 because low-level features require more computation while contributing less to deep integration models.The paper defines these three feature levels as high-level features.
- Dual-branch Global Context Module: DGCM transforms fused features into two multi-scale branches, applies MSCA to each, combines the branches, and uses a residual structure to produce enhanced features.The branches are represented by F_c and F_p, whose attention-weighted outputs are fused before residual integration.
- Dual-branch Global Context Module: The proposed DGCM adaptively extracts multi-scale information from a specific level during training to enhance ACFM fusion features.The module is designed to exploit global context within the fused features.
- Loss Function: The loss uses weighted BCE and weighted IoU components, assigning different pixel weights so hard pixels receive more attention.Weighted BCE supplies pixel restriction, while weighted IoU contributes global-structure restriction.
4 Experiments
Experiments evaluate C2F-Net on three benchmark datasets using quantitative, qualitative, and ablation studies. The results show stronger performance than prior methods and support the contributions of ACFM, DGCM, and MSCA.
- Experimental Setup: Experiments use three public COD benchmark datasets and four evaluation metrics: Sα, Eφ, Fwβ, and M.The metrics measure structural similarity, visual-perception-based accuracy, weighted precision and recall, and pixel-level error.
- Quantitative Evaluation: C2F-Net improves over ResNet50-based SINet by 4.54% in Sα, 8.71% in Eφ, and 18.35% in Fwβ on average across three datasets.The paper reports better performance on all four evaluation indicators compared with previous methods.
- Qualitative Evaluation: Qualitative comparisons show that C2F-Net detects more accurate and complete camouflaged objects with richer details than other cutting-edge models.Figure 3 examples cover aquatic, terrestrial, flying, amphibious, and other COD10K super-classes.
- Ablation Study: Adding ACFM improves the Basic configuration, while adding DGCM also improves Basic performance across the three benchmark datasets.These ablations separately support the effectiveness of attention-induced cross-level fusion and dual-branch global context modeling.
- Ablation Study: The complete Basic+ACFM+DGCM model is generally better than the other ablation settings, and visual comparisons favor its complete structure.No.4 is the complete model, whereas No.1 removes ACFM and DGCM, No.2 removes DGCM, and No.3 replaces ACFM with upsampling followed by addition.
- Ablation Study: Replacing MSCA with convolution lowers performance, while MSCA improves Fwβ by 1.9% on CAMO-Test.C2F-Net outperforms the MSCA→Conv variant on three datasets for two evaluation metrics.