CCDNet: Learning to Detect Camouflage against Distractors in Infrared Small Target Detection

Zikai Liao; Zhaozheng Yin

cs.CVMar 31, 2026

CCDNet: Learning to Detect Camouflage against Distractors in Infrared Small Target Detection

Authors: Zikai Liao, Zhaozheng Yin

Organizations: Stony Brook University

Abstract

Infrared target detection (IRSTD) tasks have critical applications in areas like wilderness rescue and maritime search. However, detecting infrared targets is challenging due to their low contrast and tendency to blend into complex backgrounds, effectively camouflaging themselves. Additionally, other objects with similar features (distractors) can cause false alarms, further degrading detection performance. To address these issues, we propose a novel Camouflage-aware Counter-Distraction Network (CCDNet) in this paper. We design a backbone with Weighted Multi-branch Perceptrons (WMPs), which aggregates self-conditioned multi-level features to accurately represent the target and background. Based on these rich features, we then propose a novel Aggregation-and-Refinement Fusion Neck (ARFN) to refine structures/semantics from shallow/deep features maps, and bidirectionally reconstruct the relations between the targets and the backgrounds, highlighting the targets while suppressing the complex backgrounds to improve detection accuracy. Furthermore, we present a new Contrastive-aided Distractor Discriminator (CaDD), enforcing adaptive similarity computation both locally and globally between the real targets and the backgrounds to more precisely discriminate distractors, so as to reduce the false alarm rate. Extensive experiments on infrared image datasets confirm that CCDNet outperforms other state-of-the-art methods.

Figures & tables

Figure 1: Illustrations of two key challenges in IRSTD tasks: target camouflage and distractor interference.

Figure 2: Overview of our proposed CCDNet. It has a backbone with proposed WMPs to extract rich contextual features, and ARFN with proposed TBSG and BOSE to highlight target features for camouflage detection via adaptive guidance from key semantics and structures. The proposed CaDD, with LCM and GCM, enables the network to better differentiate between real targets and their distractors.

Figure 3: Illustration of our proposed LCM and GCM in CaDD. $\mathbf{P}^{\text{in}}$ and $\mathbf{P}^{\text{out}}$ denote the nine-region area from the input feature map and the LCM-processed feature map. Pos. and Neg. means the positive and negative sample. Both LCM and GCM apply to all four output feature maps from the backbone. For visualization purposes, we only display its mechanism with only one feature map.

Category	Method		IRSTD-1k [ 57 ]			NUDT-SIRST [ 29 ]			NUAA-SIRST [ 4 ]
Category	Method		P ↑	R ↑	F-1 ↑	P ↑	R ↑	F-1 ↑	P ↑	R ↑	F-1 ↑
General Detectors	RetinaNet [ 39 ]	TPMAI’20	81.28	85.72	83.72	88.81	79.52	83.91	90.18	87.92	89.04
	Deformable DETR [ 62 ]	ICLR’21	73.19	62.79	67.59	79.97	75.68	77.77	80.01	71.20	75.35
	FCOS [ 43 ]	TPAMI’20	85.91	88.45	87.16	90.31	89.89	90.10	90.03	84.32	87.08
Camouflage Detectors	SINet [ 9 ]	CVPR’20	80.18	84.28	82.17	76.52	81.37	78.87	86.94	88.20	87.57
Camouflage Detectors	FEDER [ 17 ]	CVPR’23	84.93	89.29	87.06	90.47	85.75	88.05	90.33	87.78	89.04

Table 1: Quantitative results of our proposed method and other comparison methods. The best values are highlighted in bold fonts, and the second best are underlined. Our proposed method achieves the overall best results on different metrics and datasets.

Method	Params. (M) ↓	FLOPs (G) ↓	FPS ↑
ISNet [ 57 ]	1.094	122.55	37.62
DAGNet [ 13 ]	64.20	123.98	38.49
Ours	65.88	121.54	39.76
ACM [ 4 ]	0.520	8.385	42.74
IRPruneDet [ 56 ]	0.180	0.938	51.98
MSHNet [ 32 ]	4.07	33.51	69.54

Table 2: Model complexity of IRSTD methods ranked by FPS.

Figure 4: Qualitative results of our proposed method and other comparison methods. For better visualization, we only pick at least one method from each category to showcase their detection performance. The red , green , and yellow boxes are ground truths, detection results, and false alarms. Image without any red box indicates a missed detection. We put close-ups for each detection result for better visual comparison. Qualitative comparisons of the other methods can be found in the supplement. Our CCDNet yields superior results with fewer false alamrs, fewer missed detections, and better overlapped bounding boxes.

#	WMP	ARFN	CaDD	P ↑	R ↑	F1 ↑
1	-	-	-	87.68	88.09	87.89
2	$\checkmark$	-	-	88.91	89.54	89.22
3	-	$\checkmark$	-	88.33	89.25	88.79
4	-	-	$\checkmark$	89.14	89.27	89.21
5	$\checkmark$	$\checkmark$	-	89.57	90.51	89.68
6	$\checkmark$	$\checkmark$	$\checkmark$	92.08	92.64	92.36

Table 3: Ablation study of each proposed component.

#	Backbone Type	P ↑	R ↑	F-1 ↑
1	ResNet Block [ 20 ]	90.84	88.62	89.72
2	ResNet Bottleneck [ 20 ]	88.68	89.51	89.09
3	Inception Block [ 41 ]	88.92	90.17	89.54
4	RepVGG Block [ 7 ]	91.29	89.79	90.53
5	WMP (Ours)	92.08	92.64	92.36

Table 4: Ablation study on the proposed WMP.

Figure 5: Heatmaps for the outputs of last WMP from the stage 1 backbone. The results verify our WMP can accurately perceive the targets.

Figure 6: Heatmap comparison of the first stage output feature map with and without using ARFN for camouflage detection. The green bounding box indicates the target region.

Figure 7: Visual results of our proposed ARFN with BOSE and TBSG. This validates that ARFN can effectively highlight target features and suppress background features during feature fusion.

#	Fusion Type	P ↑	R ↑	F-1 ↑
1	FPN [ 31 ]	86.23	88.57	87.38
2	PANet [ 33 ]	88.96	89.08	89.02
3	BiFPN [ 42 ]	85.25	86.48	85.86
4	ACM [ 4 ]	89.46	90.63	90.05
5	Ours w/o TBSG & BOSE	83.39	84.98	84.18
6	Ours w/o TBSG	86.12	85.38	85.75

Table 5: Ablation study on our proposed TBSG and BOSE.

#	Contrastive Module		P ↑	R ↑	F-1 ↑
#	LCM	GCM	P ↑	R ↑	F-1 ↑
1	-	-	88.57	90.21	89.38
2	$\checkmark$	-	89.98	90.98	90.47
3	-	$\checkmark$	90.79	91.03	90.91
4	$\checkmark$	$\checkmark$	92.08	92.64	92.36

Table 6: Ablation study on the proposed LCM and GCM.

Figure 8: 3D Visualization of the impact of our proposed LCM.

Figure 9: Visualization of the effectiveness of the GCM. The red , green , and yellow boxes are ground truths, detection results, and false alarms.

Figure 10: Structure transformation of our proposed WMP. States (1)-(4) show the structure transformations, and steps (a)-(c) demonstrate the parameter transformation. $\mu$ , $\sigma$ , $\gamma$ , and $\beta$ denote the accumulated mean, standard deviation, and learned scaling factor and bias of the BN layer. The illustration of the learnable parameter integration is in (b).

Figure 11: Related works using background features to guide vision tasks. Our ARFN studies this idea and propose to use background features to form a target-centric feature representation favorable for detection.

Figure 12: Illustration of the data conversion process.

Figure 13: Network visual heatmap comparisons between our CCDNet without and with the use of ARFN. The results suggest that our proposed ARFN can facilitate our CCDNet focusing more on the real targets and less on the backgrounds. Details can be found in Sec. 7.2.1 .

Strategy	P ↑	R ↑	F-1 ↑
TBSG w/o concatenation	90.10	89.17	89.63
TBSG w/o channel refinement	91.38	89.06	90.21
TBSG+CBAM [ 48 ]	91.19	90.74	90.96
TBSG+ECA Block [ 46 ]	91.66	92.57	92.11
TBSG+LCDC [ 12 ]	92.10	92.43	92.26
TBSG+CR (Ours)	92.08	92.64	92.36

Table 7: Quantitative results when using different feature refinements in ARFN. Details can be found in Sec. 7.2.1 .

Figure 14: Different variants of our LCM, including whether to use the feature difference calculation (FD), multi-branch regional feature representation (MB), or the use of shared MLP (SM). Tab. 8 shows that our proposed variant D achieves the best performance. Details can be found in Sec. 7.2.2 .

Variant	Components			P ↑	R ↑	F-1 ↑
Variant	FD	MB	SM	P ↑	R ↑	F-1 ↑
1	-	-	-	88.61	90.36	89.48
2	$\checkmark$	-	-	90.69	90.45	90.57
3	$\checkmark$	$\checkmark$	-	91.29	90.81	91.05
4 (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	92.08	92.64	92.36

Table 8: Effectiveness of various components in our LCM. Details can be found in Sec. 7.2.2 .

Figure 15: Illustration and Visualization of the impact of our proposed LCM in training stage. The results show that our proposed LCM can effectively highlight the target region and suppress its surrounding regions, which can facilitate an accurate detection performance. Details can be found in Sec. 7.2.2 .

Figure 16: The structure of cross-attention GCM (CA-GCM), another adaptive global contrastive module we design for ablation study. Details can be found in Sec. 7.2.3 .

Model	P ↑	R ↑	F-1 ↑
CCDNet w/o GCM	88.89	90.98	89.92
CCDNet+GCM (Ours)	92.08	92.64	92.36
CCDNet+CA-GCM	89.74	91.66	91.20
CCDNet+ADCR [ 12 ]	89.09	91.75	90.40

Table 9: Performance of our CCDNet employed with different global contrastive modules. Details can be found in Sec. 7.2.3 .

$t_{1}$	$t_{2}$	P ↑	R ↑	F-1 ↑
0.9	0.2	91.88	92.54	92.21
0.8	0.2	92.08	92.64	92.36
0.7	0.2	89.63	92.09	90.84
0.6	0.2	87.92	90.65	89.26
0.5	0.2	86.17	90.84	88.45
0.8	0.3	91.89	91.65	91.77

Table 10: Impact of different thresholds on our GCM.

t_{1}

and

t_{2}

are two confidence thresholds introduced in Sec. 3.4.2 , respectively. Details can be found in Sec. 7.2.3 .

Figure 17: Illutrations of positive samples ( red boxes ) and negative samples ( yellow boxes ) selected in different global contrastive learning modules during network training. Details can be found in Sec. 7.2.3 .

Figure 18: Qualitative results of all the comparison methods regarding a target camouflage scenario. The red , green , and yellow boxes are ground truths, detection results, and false alarms. An image without any red box indicates a missed detection. We put close-ups for each detection result for better visual comparison.

Figure 19: Qualitative results of all the comparison methods regarding a target camouflage scenario. The red , green , and yellow boxes are ground truths, detection results, and false alarms. An image without any red box indicates a missed detection. We put close-ups for each detection result for better visual comparison.

Figure 20: Qualitative results of all the comparison methods regarding a target camouflage scenario. The red , green , and yellow boxes are ground truths, detection results, and false alarms. An image without any red box indicates a missed detection. We put close-ups for each detection result for better visual comparison.

Figure 21: Qualitative results of all the comparison methods regarding a target camouflage scenario. The red , green , and yellow boxes are ground truths, detection results, and false alarms. An image without any red box indicates a missed detection. We put close-ups for each detection result for better visual comparison.

Infrared Object Detection

Infrared object detection locates and classifies objects in thermal imagery, including small, low-contrast targets and visible–thermal scenes.

See all 33 papers

Selective Attention-Based Network for Robust Infrared Small Target Detection
Apr 27, 2026Yingming Zhang, Wuqi Su, Qing Xiao +1Infrared Small Target Detection Multi-Scale Feature Fusion
School of Software and Internet of Things Engineering, Zhejiang Gongshang University · School of Computer Science and Technology, Tiangong University
Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation
Jun 20, 2026Houzhang Fang, Ruixuan Huang, Qiuhuan Chen +3Attention-Guided Distillation Infrared Small Target Detection
Xidian University, Xi’an, China · Huazhong University of Science and Technology, Wuhan, China
SANet: Selective Attention Network for Infrared Small Target Detection
Oct 7, 2026Yingmei Zhang, Wangtao Bao, Qin Xiao +5Small Object Detection Infrared Small Target Detection
School of Software and Internet of Things Engineering, Jiangxi University of Finance and Economics, Nanchang 330032, China · School of Information Management and Mathematics, Jiangxi University of Finance and Economics, Nanchang 330032, China · School of Computer Science and Technology, Tiangong University, Tianjin 300387, China

Apr 27, 2026J/K move · Enter open · S save

Selective Attention-Based Network for Robust Infrared Small Target Detection

Yingming Zhang, Wuqi Su, Qing Xiao, Yonggang Yang

School of Software and Internet of Things Engineering, Zhejiang Gongshang University · School of Computer Science and Technology, Tiangong University

Infrared Small Target Detection Multi-Scale Feature Fusion Convolutional Neural Networks

Open paper AskPDF arXiv

Infrared small target detection (IRSTD) plays a pivotal role in a broad spectrum of mission-critical applications, including maritime surveillance, military search and rescue, early warning systems, and precision-guided strikes, all of which demand the precise identification of dim, sub-pixel targets amid highly cluttered infrared backgrounds. Despite significant progress driven by deep learning methods, fundamental challenges persist: infrared small targets occupy extremely limited spatial extents (often only a few pixels), exhibit low signal-to-clutter ratios, and are easily confused with structurally complex backgrounds that frequently induce false alarms. Existing encoder-decoder architectures suffer from two key limitations - an information bottleneck in early convolutional stages that undermines fine-grained target perception, and static skip connections that lack the dynamic adaptability required to discriminate between genuine targets and pseudo-target regions. To address these challenges, we propose SANet, a Selective Attention-based Network built upon the classical U-Net framework and augmented with two novel components: (1) a Dual-path Semantic-aware Module (DSM) that integrates standard convolutions for local spatial detail preservation with pinwheel-shaped convolutions for expanded, direction-sensitive receptive fields, followed by a Convolutional Block Attention Module (CBAM) for fine-grained spatial-channel feature recalibration; and (2) a Selective Attention Fusion Module (SAFM) that replaces conventional static skip connections with a spatially adaptive, learnable weighting mechanism to perform context-aware, cross-scale feature fusion.

Abstract

Infrared Object Detection

Explore similar work

Selective Attention-Based Network for Robust Infrared Small Target Detection

Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation

SANet: Selective Attention Network for Infrared Small Target Detection