I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection

Hongwei ZhuPeng LiHaoran XieXuefeng YanDong LiangDapeng ChenMingqiang WeiJing Qin

article2022AAAI238 citations

Proposes a boundary-guided separated attention network that mirrors human perception by decoupling foreground and background streams to locate camouflaged objects with highly ambiguous boundaries, outperforming sixteen state-of-the-art methods across standard benchmarks.

Listen

Detecting objects that visually blend into their environments is critical for high-stakes applications such as medical image segmentation, search-and-rescue operations in harsh conditions, and surveillance. Traditional computer vision methods and standard prominent-object detectors struggle in this domain because camouflaged targets actively conceal themselves, sharing colors, textures, and ambiguous boundaries with their surrounding backgrounds.

The article evaluates a human-inspired artificial intelligence architecture called the Boundary-Guided Separated Attention Network (BSA-Net). Its primary objective is to demonstrate that coordinating dual attention mechanisms—one focusing on foreground details while the other analyzes background cues—alongside targeted boundary guidance significantly enhances the detection accuracy of camouflaged objects.

The evaluated framework employs a coarse-to-fine learning strategy built on a multi-scale feature backbone. It utilizes two dedicated processing streams: a normal attention stream to isolate foreground information and a reverse attention stream that erases target interiors to evaluate the background. A specialized boundary module then integrates edge representations directly into these features. The system was validated across three benchmark datasets—CAMO (1,250 images), CHAMELEON (76 images), and the large-scale COD10K dataset (10,000 images)—and benchmarked against 16 leading baseline and state-of-the-art methods using four standard segmentation accuracy and error metrics.

The evaluation yielded several key findings. First, BSA-Net outperformed all 16 competing methods across all evaluated benchmarks, achieving the highest overall structural similarity and lowest pixel-level error. Second, on the large-scale COD10K dataset, BSA-Net demonstrated substantial gains over leading models such as SINet, improving weighted precision-recall metrics by approximately 0.148 while reducing error rates by 0.017. Third, ablation experiments confirmed that both the separated attention mechanism and the boundary guider module were essential to performance, with each contributing measurable improvements over baseline feature extraction. Finally, qualitative visual analysis verified that the network produces crisper, more complete contours and effectively identifies subtle and small-scale hidden targets that other models miss.

These findings indicate that treating foreground and background features separately provides a more robust foundation for separating targets from confusing clutter than conventional single-stream approaches. By substantially improving boundary delineation and reducing omissions, the model reduces the operational risk of missed detections in mission-critical tasks such as polyp or lung infection segmentation in healthcare, as well as rapid hazard identification during rescue missions.

Organizations developing automated image segmentation pipelines should consider adopting dual-stream foreground-background separation and explicit boundary conditioning modules to enhance edge-detection accuracy. For future development, the article recommends investigating synthetic camouflage generation to augment training pipelines and incorporating depth data alongside standard imagery to further boost detection robustness.

Confidence in these findings is supported by consistent performance across three independent benchmark datasets. However, stakeholders should note that real-world deployment will depend on operational image resolutions and environmental conditions that may differ from standard experimental datasets, warranting targeted validation before deployment in safety-critical workflows.

Hong Kong Polytechnic University Huawei Lingnan University MIIT Key Laboratory of Pattern Analysis and Machine Intelligence Nanjing University of Aeronautics and Astronautics detection algorithm object detection algorithms camouflaged object detection boundary-guided separated attention network salient object detection

Zhu et al (2022).pdf

Paper: BASNet: Boundary-Aware Salient Object Detection, Xuebin Qin et al. (2019). This paper establishes the paradigm of using dedicated boundary refinement and hybrid loss supervision to improve object delineation in dense visual segmentation.
Paper: Enhanced-alignment Measure for Binary Foreground Map Evaluation, Deng-Ping Fan et al. (2018). This paper introduces the Enhanced-alignment metric (E-measure) widely used to evaluate foreground segmentation maps against ground truth in camouflaged and salient object detection.
Paper: CBAM: Convolutional Block Attention Module, Sanghyun Woo et al. (2018). This paper provides the foundational channel and spatial attention mechanisms that inspire modular attention designs for separating foreground and background features.
Paper: Dual Attention Network for Scene Segmentation, Jun Fu et al. (2019). This work formulates dual-stream spatial and channel attention schemes essential for understanding two-stream attention pipelines in pixel-level segmentation.
Paper: Deeply Supervised Salient Object Detection with Short Connections, Qibin Hou et al. (2016). This paper demonstrates how deeply supervised skip connections and multi-level feature integration preserve fine spatial and edge details in segmentation networks.

Paper: Efficient Multi-Scale Attention Module with Cross-Spatial Learning, Daliang Ouyang et al. (2023). This paper extends multi-scale cross-spatial attention architectures to achieve lightweight, high-performance feature aggregation without reducing channel dimensions.
Paper: YOLOv12: Attention-Centric Real-Time Object Detectors, Yunjie Tian et al. (2025). This work explores attention-centric architecture design tailored for real-time detection, advancing the computational efficiency of attention modules in dense vision tasks.

Cover for I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection

Abstract

Can you find me? By simulating how humans to discover the so-called ‘perfectly’-camouflaged object, we present a novel boundary-guided separated attention network (call BSA-Net). Beyond the existing camouflaged object detection (COD) wisdom, BSA-Net utilizes two-stream separated attention modules to highlight the separator (or say the camouflaged object’s boundary) between an image’s background and foreground: the reverse attention stream helps erase the camouflaged object’s interior to focus on the background, while the normal attention stream recovers the interior and thus pay more attention to the foreground; and both streams are followed by a boundary guider module and combined to strengthen the understanding of the boundary. The core design of such separated attention is motivated by the COD procedure of humans: find the subtle difference between the foreground and background to delineate the boundary of a camouflaged object, then the boundary can help further enhance the COD accuracy. We validate on three benchmark datasets that our BSA-Net is very beneficial to detect camouflaged objects with the blurred boundaries and similar colors/patterns with their backgrounds. Extensive results exhibit very clear COD improvements on our BSA-Net over sixteen SOTAs.

Topics

detection algorithm object detection algorithms

Keywords

camouflaged object detection boundary-guided separated attention network salient object detection boundary-guided separated attention residual multi-scale feature extractor foreground attention map boundary detection feature fusion

Abstract
Introduction
Related Work
Salient Object Detection (SOD)
Camouflaged Object Detection (COD)
Methodology
Network Architecture
Residual Multi-scale Feature Extractor
Separated Attention
Boundary Guider
Loss Function
Settings
Comparison with SOTA Methods
Experiment
Ablation Study
Conclusion
Acknowledgements
References

Knowls

Knowl 1 — Boundary-Guided Separated Attention Network (BSA-Net) Architecture
model/method
BSA-Net is a coarse-to-fine deep neural network designed for camouflaged object detection (COD) that explicitly decouples foreground and background attention while incorporating boundary guidance.

Given an input image $I \in \mathbb{R}^{W \times H \times 3}$ , where $W$ and $H$ denote the width and height, feature extraction and prediction proceed as follows:
1. Backbone Feature Extraction: A Res2Net backbone extracts multi-level feature representations $F_i$ for stage $i \in \{1, 2, 3, 4, 5\}$ .
2. Context Enrichment: Multi-level features $F_2, F_3, F_4, F_5$ are processed in parallel by Residual Multi-scale Feature Extractor (RMFE) blocks to generate context-enhanced feature maps $\text{RMFE}_i$ .
3. Separated Attention and Boundary Guidance: The features are fed into Separated Attention (SEA) modules containing two streams: a reverse attention stream that erases the object interior to attend to background context, and a normal attention stream that restores the interior to focus on the foreground. Both streams are modulated by a Boundary Guider (BG) module using boundary predictions from a dedicated Boundary Detector.
4. Coarse Predictions and Refinement: Coarse prediction maps $C_i$ ( $i \in \{1, 2, 3, 4\}$ ) are obtained from the reverse attention streams under deep supervision. Shuffle Attention (SHA) blocks and feature fusion units then refine the outputs across layers, yielding four refined saliency maps $R_i$ ( $i \in \{1, 2, 3, 4\}$ ). The finest prediction $R_1$ serves as the final binary segmentation output during inference.
Knowl 2 — Separated Attention (SEA) Module
model/method
The Separated Attention (SEA) module coordinates foreground and background representations to isolate the camouflaged object boundary.

Given the $i$ -th layer multi-scale feature $\text{RMFE}_i \in \mathbb{R}^{C \times H_i \times W_i}$ and the upsampled coarse prediction from the $(i+1)$ -th deeper layer $C_{i+1}$ , the foreground and background attention maps are defined as:

$W_{\text{fai}} = \sigma(C_{i+1})$ $W_{\text{bai}} = 1 - \sigma(C_{i+1})$

where $\sigma(\cdot)$ is the sigmoid function, and both attention maps are channel-expanded to 64 channels. The background-focused and foreground-focused stream features are computed as:

$B_{ai} = \text{Out}_i = \text{Conv}_s(\text{RMFE}_i \otimes \text{expand}(W_{\text{bai}}))$ $F_{ai} = \text{Conv}_s(\text{RMFE}_i \otimes \text{expand}(W_{\text{fai}}))$

where $\text{Conv}_s$ denotes a $1 \times 1$ convolution, $\otimes$ is element-wise multiplication, and $\text{Out}_i$ represents the coarse output map supervised by ground truth. To balance the contributions of the two streams across spatial and channel dimensions, a Multi-scale Channel Attention Module (MS-CAM) computes a scale weight matrix:

$W(X) = G(\sigma(G(X))) + G(\sigma(L(X)))$

where $G(X)$ applies global average pooling followed by convolution, and $L(X)$ applies pointwise convolution for local feature extraction. Using $X = B_{ai} + F_{ai}$ , the two streams are gated, modulated by the Boundary Guider module ( $\text{BG}_i$ ), and integrated:

$\text{SEA}_{Fi} = \text{BG}_i(W(B_{ai} + F_{ai}) \otimes B_{ai}, BM)$ $\text{SEA}_{Bi} = \text{BG}_i((1 - W(B_{ai} + F_{ai})) \otimes F_{ai}, BM)$ $\text{SEA}_i = \text{SEA}_{Fi} \oplus \text{SEA}_{Bi}, \quad i \in \{2, 3, 4\}$

where $BM$ is the predicted boundary map and $\oplus$ denotes element-wise addition.
Knowl 3 — Boundary Guider (BG) Module and Boundary Detector
model/method
The Boundary Guider (BG) module injects boundary priors directly into the attention stream feature maps using conditional batch normalization / adaptive space normalization.

Given an Attention Stream Map $\text{ASM}_i$ from the SEA module and the predicted boundary map $BM$ , the boundary-guided feature map $\text{BGM}_i$ is computed as:

$\text{BGM}_i = \text{CB}(\text{ASM}_i) \otimes \gamma(BM) \oplus \beta(BM)$

where $\text{CB}$ represents a $3 \times 3$ convolution followed by batch normalization; $\gamma(BM)$ and $\beta(BM)$ are affine transformation parameters generated by applying two separate $3 \times 3$ convolutional layers to $BM$ , projecting the single-channel boundary map to 64 channels to match the feature channels of $\text{CB}(\text{ASM}_i)$ ; $\otimes$ and $\oplus$ denote element-wise multiplication and addition, respectively.

The boundary prediction map $BM$ is produced by a dedicated boundary detector network that concatenates four multi-level feature maps from the backbone network and passes them through a convolutional layer supervised by the ground-truth boundary map $BG$ (computed as the spatial gradient map of the binary ground-truth mask).
Knowl 4 — Residual Multi-scale Feature Extractor (RMFE)
model/method
The Residual Multi-scale Feature Extractor (RMFE) expands the receptive field of backbone features $F_i$ ( $i \in \{2, 3, 4, 5\}$ ) using four cascaded residual branches.

For an input feature map $F_i$ , branch outputs $\text{Bout}_k^i$ for $k \in \{1, 2, 3, 4\}$ are computed as:

$\text{Bout}_k^i = \begin{cases} \text{Conv}_r(F_i) & k = 1 \\ \text{Conv}_r(F_i \oplus \text{Bout}_{k-1}^i) & k = 2, 3, 4 \end{cases}$

where $\oplus$ denotes element-wise addition, and $\text{Conv}_r(\cdot)$ denotes a composite block consisting of a $1 \times 1$ convolution (to reduce channel dimensionality) followed by asymmetric $1 \times 3$ and $3 \times 1$ convolutions to enlarge receptive fields while reducing computational cost.

The four branch outputs are concatenated, projected to 64 channels via a $1 \times 1$ convolution, and added to the projected input feature:

$\text{RMFE}_i = \text{Conv}(F_i) \oplus \text{Conv}\left(\text{Cat}_{k=1}^4(\text{Bout}_k^i)\right)$

where $\text{Conv}(\cdot)$ denotes a $1 \times 1$ convolution and $\text{Cat}_{k=1}^4$ denotes concatenation along the channel dimension.
Knowl 5 — Multi-Task Supervision Loss Function for BSA-Net
equation
BSA-Net is trained end-to-end using a multi-task loss that supervises 9 target maps: 4 coarse segmentation maps ( $C_1, C_2, C_3, C_4$ ), 4 refined segmentation maps ( $R_1, R_2, R_3, R_4$ ), and 1 boundary map ( $B$ ):

$\mathcal{L} = \sum_{i=1}^4 \left[ \mathcal{L}_t(C_i, G) + \mathcal{L}_t(R_i, G) \right] + \mathcal{L}_{\text{bce}}(B, BG)$

where $G$ is the binary ground-truth segmentation mask, $BG$ is the ground-truth boundary mask, and $\mathcal{L}_{\text{bce}}$ is standard binary cross-entropy loss.

For each segmentation map prediction $P$ , the combined segmentation loss $\mathcal{L}_t(P, G)$ is:

$\mathcal{L}_t = \mathcal{L}_{\text{Wbce}} + \mathcal{L}_{\text{IOU}}$

where $\mathcal{L}_{\text{IOU}}$ is the intersection-over-union loss, and $\mathcal{L}_{\text{Wbce}}$ is a weighted binary cross-entropy loss that assigns higher weights to hard-to-predict pixels:

$\mathcal{L}_{\text{Wbce}} = -\sum_{n=1}^N w_n \left[ G_n \ln(P_n) + (1 - G_n) \ln(1 - P_n) \right]$

with pixel weight $w_n = \sigma(|P_n - G_n|)$ , where $N$ is the total number of image pixels, and $P_n, G_n$ denote predicted and ground-truth values at pixel $n$ .

Knowl 6 — Comparative Performance on COD Benchmark Datasets

data/table

BSA-Net was evaluated against 16 state-of-the-art salient object detection (SOD) and camouflaged object detection (COD) methods across three benchmark datasets: CAMO, CHAMELEON, and COD10K. The evaluation metrics are Structure-measure ( $S_\alpha \uparrow$ ), Enhanced-alignment measure ( $E_\phi \uparrow$ ), Weighted F-measure ( $F_\beta^\omega \uparrow$ ), and Mean Absolute Error ( $\text{MAE} \downarrow$ ).

Methods	CAMO				CHAMELEON				COD10K
	$S_\alpha \uparrow$	$E_\phi \uparrow$	$F_\beta^\omega \uparrow$	$\text{MAE} \downarrow$	$S_\alpha \uparrow$	$E_\phi \uparrow$	$F_\beta^\omega \uparrow$	$\text{MAE} \downarrow$	$S_\alpha \uparrow$	$E_\phi \uparrow$	$F_\beta^\omega \uparrow$	$\text{MAE} \downarrow$
UNet++ (2018)	0.599	0.653	0.392	0.149	0.695	0.762	0.501	0.094	0.623	0.672	0.350	0.086
PiCANet (2018)	0.609	0.584	0.356	0.156	0.769	0.749	0.536	0.085	0.649	0.643	0.322	0.090
MSRCNN (2019)	0.617	0.669	0.454	0.133	0.637	0.686	0.443	0.091	0.641	0.706	0.419	0.073
BASNet (2019)	0.618	0.661	0.413	0.159	0.687	0.721	0.474	0.118	0.634	0.678	0.365	0.105
PFANet (2019)	0.659	0.622	0.391	0.172	0.679	0.648	0.378	0.144	0.636	0.618	0.286	0.128
CPD (2019)	0.726	0.729	0.550	0.115	0.853	0.866	0.706	0.052	0.747	0.770	0.508	0.059
HTC (2019)	0.476	0.442	0.174	0.172	0.517	0.489	0.204	0.129	0.548	0.520	0.221	0.088
EGNet (2019)	0.732	0.768	0.583	0.104	0.848	0.870	0.702	0.050	0.737	0.779	0.509	0.056
ANet-SRM (2019)	0.682	0.685	0.484	0.126	-	-	-	-	-	-	-	-
SINet (2020)	0.751	0.771	0.606	0.100	0.869	0.891	0.740	0.044	0.771	0.806	0.551	0.051
PraNet (2020)	0.769	0.824	0.663	0.094	0.860	0.907	0.763	0.044	0.789	0.861	0.629	0.045
MCIF-Net (2021)	0.784	0.845	0.677	0.084	-	-	-	-	0.787	0.872	0.636	0.042
TANet (2021)	0.793	0.834	0.690	0.083	0.888	0.911	0.786	0.036	0.803	0.848	0.629	0.041
R-MGL (2021)	0.775	0.847	0.673	0.088	0.893	0.923	0.813	0.030	0.814	0.865	0.666	0.035
PFNet (2021)	0.782	0.841	0.695	0.085	0.882	0.931	0.810	0.033	0.800	0.877	0.660	0.040
Rank-Net (2021)	0.787	0.838	0.696	0.080	0.890	0.935	0.822	0.030	0.804	0.880	0.673	0.037
BSA-Net (Ours)	0.796	0.851	0.717	0.079	0.895	0.946	0.841	0.027	0.818	0.891	0.699	0.034

BSA-Net outperforms all sixteen comparison models on all four metrics across CAMO, CHAMELEON, and COD10K datasets.

Knowl 7 — Ablation Study on BSA-Net Modules

data/table

An ablation study evaluated the progressive contributions of the Separated Attention (SEA) and Boundary Guider (BG) modules starting from a baseline composed of the Res2Net backbone and Residual Multi-scale Feature Extractor (RMFE).

Model	CAMO-Test				CHAMELEON				COD10K-Test
	$S_\alpha \uparrow$	$E_\phi \uparrow$	$F_\beta^\omega \uparrow$	$\text{MAE} \downarrow$	$S_\alpha \uparrow$	$E_\phi \uparrow$	$F_\beta^\omega \uparrow$	$\text{MAE} \downarrow$	$S_\alpha \uparrow$	$E_\phi \uparrow$	$F_\beta^\omega \uparrow$	$\text{MAE} \downarrow$
a. Baseline	0.783	0.840	0.691	0.083	0.879	0.933	0.807	0.032	0.805	0.880	0.672	0.037
b. Baseline+SEA	0.796	0.854	0.716	0.080	0.882	0.934	0.815	0.030	0.811	0.881	0.685	0.037
c. Baseline+BG	0.786	0.848	0.704	0.081	0.875	0.921	0.808	0.031	0.812	0.883	0.686	0.036
d. Ours (Full BSA-Net)	0.796	0.851	0.717	0.079	0.895	0.946	0.841	0.027	0.818	0.891	0.699	0.034

Adding SEA (Model b vs. Model a) consistently boosts $F_\beta^\omega$ from 0.691 to 0.716 on CAMO-Test, 0.807 to 0.815 on CHAMELEON, and 0.672 to 0.685 on COD10K-Test, confirming that separating foreground and background attention streams helps resolve ambiguity at camouflage boundaries.
Adding BG alone (Model c vs. Model a) also improves boundary sensitivity.
The full model combining both SEA and BG (Model d) achieves the best overall performance across all three datasets.

Knowl 8 — Experimental Setup and Training Configuration for BSA-Net
experimental setup
BSA-Net is implemented in PyTorch and evaluated under the following settings:
- Benchmark Datasets:
  - CAMO: 1250 camouflaged images (1000 train, 250 test) and 1250 non-camouflaged images from MS-COCO (1000 train, 250 test).
  - CHAMELEON: 76 images used exclusively for testing.
  - COD10K: 6000 training images and 4000 test images.
- Network Initialization: Res2Net (blocks 1 to 5) pretrained on ImageNet serves as the backbone. Convolutional and linear layers are initialized with Kaiming normal initialization.
- Optimization: Adam optimizer with an initial learning rate of $8 \times 10^{-5}$ and weight decay of $0.1$ .
- Training Parameters: Batch size of 36, trained for up to 35 epochs. Input images are resized to $384 \times 384$ using bilinear interpolation.
- Hardware: Single Nvidia GeForce RTX 3090 GPU (24GB memory).

Coverage note — None was omitted; all key architectural components (RMFE, SEA, BG, Boundary Detector), multi-task loss formulation, benchmark evaluations, ablation studies, and implementation details are fully captured.

References

1.Bhajantri, N. U.; and Nagabhushan, P. 2006. Camouflage defect identification: a novel approach. In 9th International Conference on Information Technology (ICIT’06), 145–148. IEEE.
2.Chen, K.; Pang, J.; Wang, J.; Xiong, Y.; Li, X.; Sun, S.; Feng, W.; Liu, Z.; Shi, J.; Ouyang, W.; et al. 2019. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4974–4983.
3.Chen, S.; Tan, X.; Wang, B.; and Hu, X. 2018. Reverse attention for salient object detection. In Proceedings of the European Conference on Computer Vision (ECCV), 234–250.
4.Chu, H.; Hsu, W.; Mitra, N. J.; Cohen-Or, D.; Wong, T.; and Lee, T. 2010. Camouflage images. ACM Trans. Graph., 29(4): 51:1–51:8.
5.Dai, Y.; Gieseke, F.; Oehmcke, S.; Wu, Y.; and Barnard, K. 2021. Attentional Feature Fusion. In IEEE Winter Conference on Applications of Computer Vision, WACV 2021, Waikoloa, HI, USA, January 3-8, 2021, 3559–3568. IEEE.
6.Dong, B.; Zhuge, M.; Wang, Y.; Bi, H.; and Chen, G. 2021. Towards Accurate Camouflaged Object Detection with Mixture Convolution and Interactive Fusion. CoRR, abs/2101.05687.
7.Fan, D.; Gong, C.; Cao, Y.; Ren, B.; Cheng, M.; and Borji, A. 2018. Enhanced-alignment Measure for Binary Foreground Map Evaluation. In Lang, J., ed., Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden, 698–704. ijcai.org.
8.Fan, D.-P.; Cheng, M.-M.; Liu, Y.; Li, T.; and Borji, A. 2017. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, 4548–4557.
9.Fan, D.-P.; Ji, G.-P.; Sun, G.; Cheng, M.-M.; Shen, J.; and Shao, L. 2020a. Camouflaged object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2777–2787.
10.Fan, D.-P.; Ji, G.-P.; Zhou, T.; Chen, G.; Fu, H.; Shen, J.; and Shao, L. 2020b. Pranet: Parallel reverse attention network for polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 263–273. Springer.
11.Feng, M.; Lu, H.; and Ding, E. 2019. Attentive feedback network for boundary-aware salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1623–1632.
12.Gao, S.; Cheng, M.; Zhao, K.; Zhang, X.; Yang, M.; and Torr, P. H. S. 2021. Res2Net: A New Multi-Scale Backbone Architecture. IEEE Trans. Pattern Anal. Mach. Intell., 43(2): 652–662.
13.Huang, Z.; Huang, L.; Gong, Y.; Huang, C.; and Wang, X. 2019. Mask scoring r-cnn. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6409–6418.
14.Huerta, I.; Rowe, D.; Mozerov, M.; and Gonzalez, J. 2007. Improving background subtraction based on a casuistry of colour-motion segmentation problems. In Iberian Conference on Pattern Recognition and Image Analysis, 475–482. Springer.
15.Kavitha, C.; Rao, B. P.; and Govardhan, A. 2011. An efficient content based image retrieval using color and texture of image sub blocks. International Journal of Engineering Science and Technology (IJEST), 3(2): 1060–1068.
16.Le, T.; Nguyen, T. V.; Nie, Z.; Tran, M.; and Sugimoto, A. 2019. Anabranch network for camouflaged object segmentation. Comput. Vis. Image Underst., 184: 45–56.
17.Li, A.; Zhang, J.; Lv, Y.; Liu, B.; Zhang, T.; and Dai, Y. 2021. Uncertainty-Aware Joint Salient Object and Camouflaged Object Detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, 10071–10081. Computer Vision Foundation / IEEE.
18.Liu, J.; Zhang, J.; and Barnes, N. 2021. Confidence-Aware Learning for Camouflaged Object Detection. CoRR, abs/2106.11641.
19.Liu, N.; Han, J.; and Yang, M.-H. 2018. Picanet: Learning pixel-wise contextual attention for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 3089–3098.
20.Lv, Y.; Zhang, J.; Dai, Y.; Li, A.; Liu, B.; Barnes, N.; and Fan, D. 2021. Simultaneously Localize, Segment and Rank the Camouflaged Objects. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, 11591–11601. Computer Vision Foundation / IEEE.
21.Mao, Y.; Zhang, J.; Wan, Z.; Dai, Y.; Li, A.; Lv, Y.; Tian, X.; Fan, D.; and Barnes, N. 2021. Transformer Transforms Salient Object Detection and Camouflaged Object Detection. CoRR, abs/2104.10127.
22.Margolin, R.; Zelnik-Manor, L.; and Tal, A. 2014. How to evaluate foreground maps? In Proceedings of the IEEE conference on computer vision and pattern recognition, 248–255.
23.Mei, H.; Ji, G.; Wei, Z.; Yang, X.; Wei, X.; and Fan, D. 2021. Camouflaged Object Segmentation With Distraction Mining. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, 8772–8781. Computer Vision Foundation / IEEE.
24.Pang, Y.; Zhao, X.; Zhang, L.; and Lu, H. 2020. Multi-scale interactive network for salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9413–9422.
25.Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E.; DeVito, Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B.; Fang, L.; Bai, J.; and Chintala, S. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Wallach, H. M.; Larochelle, H.; Beygelzimer, A.; d’Alche-Buc, F.; Fox, E. B.; and Garnett, R., eds., Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, 8024–8035.
26.Perazzi, F.; Krähenbühl, P.; Pritch, Y.; and Hornung, A. 2012. Saliency filters: Contrast based filtering for salient region detection. In 2012 IEEE conference on computer vision and pattern recognition, 733–740. IEEE.
27.Qin, X.; Zhang, Z.; Huang, C.; Gao, C.; Dehghan, M.; and Jagersand, M. 2019. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7479–7489.
28.Ren, J.; Hu, X.; Zhu, L.; Xu, X.; Xu, Y.; Wang, W.; Deng, Z.; and Heng, P. 2021. Deep Texture-Aware Features for Camouflaged Object Detection. CoRR, abs/2102.02996.
29.Skurowski, P.; Abdulameer, H.; Błaszczyk, J.; Depta, T.; Kornacki, A.; and Kozieł, P. 2018. Animal camouflage analysis: Chameleon database. Unpublished Manuscript.
30.Sun, Y.; Chen, G.; Zhou, T.; Zhang, Y.; and Liu, N. 2021. Context-aware Cross-level Fusion Network for Camouflaged Object Detection. In Zhou, Z., ed., Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada, 19-27 August 2021, 1025–1031. ijcai.org.
31.Tankus, A.; and Yeshurun, Y. 2001. Convexity-Based Visual Camouflage Breaking. Comput. Vis. Image Underst., 82(3): 208–237.
32.Wei, J.; Wang, S.; and Huang, Q. 2020. F3Net: Fusion, Feedback and Focus for Salient Object Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, 12321–12328.
33.Wei, J.; Wang, S.; Wu, Z.; Su, C.; Huang, Q.; and Tian, Q. 2020. Label Decoupling Framework for Salient Object Detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 13022–13031. IEEE.
34.Wu, Y.; Gao, S.; Mei, J.; Xu, J.; Fan, D.; Zhang, R.; and Cheng, M. 2021. JCS: An Explainable COVID-19 Diagnosis System by Joint Classification and Segmentation. IEEE Trans. Image Process., 30: 3113–3126.
35.Wu, Z.; Su, L.; and Huang, Q. 2019a. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3907–3916.
36.Wu, Z.; Su, L.; and Huang, Q. 2019b. Stacked cross refinement network for edge-aware salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 7264–7273.
37.Xue, F.; Guoying, C.; Hong, R.; and Gu, J. 2015. Camouflage texture evaluation using a saliency map. Multim. Syst., 21(2): 169–175.
38.Yan, J.; Le, T.; Nguyen, K.; Tran, M.; Do, T.; and Nguyen, T. V. 2021. MirrorNet: Bio-Inspired Camouflaged Object Segmentation. IEEE Access, 9: 43290–43300.
39.Yang, Q.-L. Z. Y.-B. 2021. SA-Net: Shuffle Attention for Deep Convolutional Neural Networks. arXiv preprint arXiv:2102.00240.
40.Zhai, Q.; Li, X.; Yang, F.; Chen, C.; Cheng, H.; and Fan, D. 2021. Mutual Graph Learning for Camouflaged Object Detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, 12997–13007. Computer Vision Foundation / IEEE.
41.Zhang, J.; Lv, Y.; Xiang, M.; Li, A.; Dai, Y.; and Zhong, Y. 2021. Depth-Guided Camouflaged Object Detection. CoRR, abs/2106.13217.
42.Zhao, J.-X.; Liu, J.-J.; Fan, D.-P.; Cao, Y.; Yang, J.; and Cheng, M.-M. 2019. EGNet: Edge guidance network for salient object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 8779–8788.
43.Zhao, T.; and Wu, X. 2019. Pyramid feature attention network for saliency detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3085–3094.
44.Zhou, Z.; Siddiquee, M. M. R.; Tajbakhsh, N.; and Liang, J. 2018. Unet++: A nested u-net architecture for medical image segmentation. In Deep learning in medical image analysis and multimodal learning for clinical decision support, 3–11. Springer.
45.Zhu, L.; Chen, J.; Hu, X.; Fu, C.-W.; Xu, X.; Qin, J.; and Heng, P.-A. 2019. Aggregating attentional dilated features for salient object detection. IEEE Transactions on Circuits and Systems for Video Technology, 30(10): 3358–3371.

Citation

MLA

Zhu, H., et al. “I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection”. Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 3, 2022, pp. 3608–16, https://doi.org/10.1609/AAAI.V36I3.20273.

APA

Zhu, H., Li, P., Xie, H., Yan, X., Liang, D., Chen, D., Wei, M., & Qin, J. (2022). I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3), 3608–3616. https://doi.org/10.1609/AAAI.V36I3.20273

Chicago

Zhu, H., P. Li, H. Xie, et al. 2022. “I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection”. Proceedings of the AAAI Conference on Artificial Intelligence 36 (3): 3608–16. https://doi.org/10.1609/AAAI.V36I3.20273.

Harvard

Zhu, H. et al. (2022) “I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection”, Proceedings of the AAAI Conference on Artificial Intelligence, 36(3), pp. 3608–3616. Available at: https://doi.org/10.1609/AAAI.V36I3.20273.

Vancouver

1. Zhu H, Li P, Xie H, Yan X, Liang D, Chen D, Wei M, Qin J (2022) I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection. Proceedings of the AAAI Conference on Artificial Intelligence 36:3608–3616

BibTeX

@article{Zhu_2022, title={I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection}, volume={36}, ISSN={2159-5399}, url={http://dx.doi.org/10.1609/AAAI.V36I3.20273}, DOI={10.1609/aaai.v36i3.20273}, number={3}, journal={Proceedings of the AAAI Conference on Artificial Intelligence}, publisher={Association for the Advancement of Artificial Intelligence (AAAI)}, author={Zhu, Hongwei and Li, Peng and Xie, Haoran and Yan, Xuefeng and Liang, Dong and Chen, Dapeng and Wei, Mingqiang and Qin, Jing}, year={2022}, month=June, pages={3608–3616} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

I Can Find You! Boundary-Guided Separated Attention Network for Camouflaged Object Detection

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Boundary-Guided Separated Attention Network (BSA-Net) Architecture

Knowl 2 — Separated Attention (SEA) Module

Knowl 3 — Boundary Guider (BG) Module and Boundary Detector

Knowl 4 — Residual Multi-scale Feature Extractor (RMFE)