RDANet: Relative Degradation Aware Network for Infrared Small Target Detection

Rui Liu; Jing Nie; Ying Fu

cs.CVAug 21, 2026

RDANet: Relative Degradation Aware Network for Infrared Small Target Detection

Authors: Rui Liu, Jing Nie, Ying Fu

Organizations: School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081, China · School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China

Abstract

Infrared small target detection is still challenging in remote sensing imagery, because the targets are extremely small, exhibit weak local contrast, and are often embedded in complex and highly variable backgrounds. In addition to these inherent difficulties, we observe that existing detectors often show unstable performance when the target scale changes or when the scene background varies. This scale- and scene-sensitive degradation indicates that current methods are insufficient in simultaneously preserving target structure during feature downsampling and maintaining discriminative local contrast under background shifts, which finally results in unbalanced detection performance across different conditions. To improve detection robustness, this paper proposes a Relative Degradation Aware Network (RDANet) for infrared small target detection. RDANet consists of two dedicated modules: Multi-Scale Anti-Alias Downsampling (MSAD) and Prototype-Guided Skip Memory (PGSM). MSAD introduces multi-scale anti-alias filtering together with pixel-fold aggregation to reduce aliasing effects during resolution reduction, so that target shape information can be better preserved while irrelevant background responses are suppressed. PGSM further enhances the skip features by retrieving patch-level prototypes from a shared memory and adaptively integrating them into the current representation, which helps maintain stable local contrast cues under diverse scene backgrounds. Experiments on three public benchmarks show that RDANet achieves the best performance on most evaluation metrics, while scale- and background-stratified evaluations indicate more stable behavior across target sizes and scene complexity. The code is available at https://github.com/BIT-RuiLiu/RDANet.

Figures & tables

Fig. 1: Scale-wise distribution and detection performance across target sizes on IRSTD-1k. The right panel shows the distribution of training and test samples across target size intervals. The left panel presents the IoU of representative IRSTD methods across six target size ranges. Most methods exhibit unstable or degraded performance as target size increases. In contrast, our RDANet maintains stable performance across scales.

Fig. 2: Detection results of representative deep-learning-based methods for targets at different scales and backgrounds. Changing the surrounding background causes pronounced prediction variations for enlarged targets, whereas the responses become more stable as the target size decreases.

Fig. 3: Overall architecture of the proposed RDANet. MSAD performs anti-aliased downsampling to preserve target structure and suppress background leakage, while PGSM enhances skip connections to stabilize local contrast under scene variations.

Method	Venue	IRSTD-1k			NUDT-SIRST			NUAA-SIRST
Method	Venue	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$
Top-Hat [ 12 ]	OE’96	10.06	75.11	1432	20.72	78.41	166.70	1.51	79.74	16456
WSLCM [ 14 ]	GRSL’20	3.45	72.44	6619	2.28	56.82	1309	6.39	88.74	4462
TLLCM [ 15 ]	GRSL’19	3.31	77.39	6738	2.28	62.01	1608	4.24	88.37	6243
IPI [ 16 ]	TIP’22	27.92	81.37	16.18	17.76	74.49	41.23	1.09	87.05	30467
NRAM [ 63 ]	RS’18	15.25	70.68	16.93	6.93	56.40	19.27	13.54	60.04	25.23

TABLE I: Quantitative comparison of different methods in terms of IoU (%),

{\rm P_{d}}

(%), and

{\rm F_{a}}

(

10^{-6}

). The best and second-best results are highlighted in bold and underlined , respectively.

Method	[1,15]	[16,31]	[32,63]	[64,127]	IoU $\uparrow$
MSHNet [ 10 ]	87.59	67.87	84.34	83.71	80.55
DNANet [ 11 ]	92.46	93.13	94.97	94.41	94.19
RPCANet [ 8 ]	89.15	85.27	89.69	91.95	89.37
MSDANet [ 9 ]	96.34	91.74	93.72	94.57	93.81
RDANet (Ours)	94.13	93.88	95.94	96.34	95.43

TABLE II: Scale-wise IoU(%) comparison on NUDT-SIRST.

Group	Variant	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	Params	FLOPs
Overall	Full RDANet	73.82	93.60	7.67	4.17	25.24
	w/o PGSM	71.32	92.86	11.26	4.13	25.12
	w/o MSAD	69.32	92.59	12.56	4.10	24.58
	w/o MSAD, PGSM	66.93	91.56	20.14	4.07	24.46
MSAD	w/o PixelUnshuffle	72.40	92.59	11.17	4.17	25.23
MSAD	w/o SE gate	73.10	92.93	10.74	4.16	25.24

TABLE III: Ablation study on RDANet. The top block reports the overall contribution of MSAD and PGSM, while the bottom blocks show fine-grained ablations of the two modules.

Fig. 4: Visual comparisons of multiple methods are presented. Targets that are correctly detected are highlighted with red boxes, missed targets are marked with blue boxes, and false alarms are enclosed in yellow boxes. Furthermore, close-up views of the targets are shown in the corners of the images.

Fig. 5: Frequency-spectrum visualization before and after MSAD. MSAD suppresses peripheral high-frequency components while preserving the dominant low-frequency structure.

Fig. 6: Feature visualizations of different downsampling methods across encoder stages. MSAD preserves clearer and more compact target responses, effectively maintaining structural integrity.

Fig. 7: Visualization of feature maps before and after PGSM enhancement. PGSM significantly amplifies target activations and highlights clearer target boundaries, providing more distinctive and shape-consistent representations.

Method	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	Params	FLOPs
w/ Pooling	66.93	91.56	20.14	4.07	24.46
w/ Stride Conv	67.10	91.58	23.22	4.26	25.06
w/ SPDConv [ 52 ]	67.87	92.25	15.63	5.07	29.41
w/ MSAD	69.32	92.59	12.56	4.10	24.58

TABLE IV: Ablation study on different downsampling methods.

Method	$[1,15]$	$[16,31]$	$[32,63]$	$[64,127]$	$[128,255]$	$[256,\infty)$	SB-IoU $\uparrow$	Avg. Rank $\downarrow$
DNANet [ 11 ]	40.18	48.52	60.24	67.62	74.25	85.28	62.68	4.17
MSHNet [ 10 ]	43.63	56.81	61.43	65.39	77.96	83.57	64.80	3.67
RPCANet [ 8 ]	50.00	57.40	60.61	73.82	75.61	51.62	61.51	3.50
MSDANet [ 9 ]	50.45	68.71	66.56	71.04	81.86	77.26	69.31	2.00
RDANet (Ours)	56.05	66.76	66.08	75.78	80.04	84.91	71.60	1.67

TABLE V: Scale-stratified IoU(%) comparison on IRSTD-1k. SB-IoU denotes the average IoU over scale bins, and Avg. Rank denotes the average method rank across bins.

Method	Low	Medium	High	Mean $\uparrow$	BG-Gap $\downarrow$
DNANet [ 11 ]	66.00	66.96	62.20	65.71	4.76
MSHNet [ 10 ]	70.87	69.79	64.44	67.87	6.43
RPCANet [ 8 ]	50.13	70.87	67.14	63.50	20.74
MSDANet [ 9 ]	69.19	75.78	70.84	71.89	6.59
RDANet (Ours)	75.92	74.27	71.96	73.82	3.97

TABLE VI: Background-stratified IoU(%) comparison on IRSTD-1k. BG-Gap is the gap between the best and worst background-complexity groups.

Variant	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	Params	FLOPs
Direct skip	71.32	92.86	11.26	4.09	25.12
SE-enhanced skip	72.31	93.27	16.97	4.10	25.13
CBAM-enhanced skip	70.92	92.93	30.38	4.10	25.16
Prototype memory	70.64	93.27	24.50	4.16	25.24
Frequency memory w/o shift	72.60	92.93	11.86	4.17	25.24
PGSM (full)	73.82	93.60	7.67	4.17	25.24

TABLE VII: Comparison of PGSM with alternative skip-refinement mechanisms on IRSTD-1k. Params and FLOPs are measured in M and G, respectively.

Kernels used	Fusion type	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$
$K{=}\{3\}$	–	72.82	92.93	12.85
$K{=}\{5\}$	–	72.10	94.61	16.30
$K{=}\{7\}$	–	72.82	92.59	9.11
$K{=}\{3,5\}$	Adaptive	73.28	92.59	16.34
$K{=}\{3,7\}$	Adaptive	72.12	93.60	13.61
$K{=}\{5,7\}$	Adaptive	71.74	91.25	13.95

TABLE VIII: Ablation study on blur-kernel configurations and fusion strategies in MSAD on IRSTD-1k.

Method	IRSTD-1k $\rightarrow$ NUAA-SIRST			NUAA-SIRST $\rightarrow$ IRSTD-1k			Efficiency
Method	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	Params	FLOPs	FPS
DNANet [ 11 ]	56.79	93.58	101.49	47.70	85.71	125.94	4.70	57.13	6
MSHNet [ 10 ]	66.65	92.66	22.17	54.54	85.71	61.64	4.06	24.43	20
RPCANet [ 8 ]	40.88	75.23	100.60	30.78	87.63	315.73	0.68	178.28	23
MSDANet [ 9 ]	63.90	92.70	17.97	52.70	87.54	101.29	4.79	84.90	16
RDANet (Ours)	71.83	98.17	17.43	60.74	90.91	60.24	4.17	25.24	15

TABLE IX: Cross-dataset generalization and efficiency under two transfer settings.

Method	Pixel interval						IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$
Method	[1,15]	[16,31]	[32,63]	[64,127]	[128,255]	[256, $\infty$ )	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$
DNANet [ 11 ]	37.66 (-2.52)	52.95 (+4.43)	58.49 (-1.75)	63.32 (-4.30)	72.63 (-1.62)	84.27 (-1.01)	63.74 (-1.97)	89.12 (-2.72)	19.59 (+1.98)
MSHNet [ 10 ]	44.56 (+0.93)	57.96 (+1.15)	60.07 (-1.36)	64.60 (-0.79)	76.38 (-1.58)	85.64 (+2.07)	67.56 (-0.31)	93.88 (+1.02)	9.49 (+0.61)
RPCANet [ 8 ]	35.85 (-14.15)	51.92 (-5.48)	58.77 (-1.84)	72.18 (-1.64)	73.51 (-2.10)	52.15 (+0.53)	60.37 (-3.13)	88.66 (+1.03)	20.80 (+12.15)
MSDANet [ 9 ]	54.91 (+4.46)	64.06 (-4.65)	65.20 (-1.36)	69.90 (-1.14)	78.66 (-3.20)	83.72 (+6.46)	71.44 (-0.45)	88.89 (-5.39)	8.62 (-2.77)
RDANet (Ours)	56.73 (+0.68)	64.60 (-2.16)	65.10 (-0.98)	75.96 (+0.18)	79.41 (-0.63)	85.11 (+0.20)	73.12 (-0.70)	91.58 (-2.02)	6.28 (-1.39)

TABLE X: Re-training results on the augmented IRSTD-1k dataset. The gray values in parentheses indicate the change relative to the original result in the same metric/bin.

PGSM Blocks	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$	Params	FLOPs
0	66.93	91.56	20.14	4.07	24.46
1	69.14	92.86	15.35	4.07	24.70
2	70.22	91.58	10.24	4.08	24.84
3	70.96	92.25	12.43	4.09	24.98
4	71.32	92.86	11.26	4.13	25.12

TABLE XI: Ablation study on the number of PGSM blocks.

Setting	Variant	IoU $\uparrow$	${\rm P_{d}}\uparrow$	${\rm F_{a}}\downarrow$
Memory size	$[64,32,16,8]$	70.93	91.58	11.12
Memory size	$[128,64,32,16]$	72.65	93.27	13.40
Memory size	$[256,128,64,32]$ (default)	73.82	93.60	7.67
Memory size	$[512,256,128,64]$	71.46	92.26	9.75
Top- $K$ retrieval	$[4,4,4,2]$	70.86	93.60	17.82
Top- $K$ retrieval	$[8,8,8,4]$ (default)	73.82	93.60	7.67

TABLE XII: Sensitivity analysis of PGSM on IRSTD-1k. The default configuration is highlighted in gray.

Fig. 8: ROC curves on the IRSTD-1k dataset.

Explore similar work

Degraded Infrared Small Object Detection via Degradation-Adapted Physics-Guided Restoration
Aug 10, 2026Xinkai Lu, Wenjun Chen, Yi Li +2Image Corruption Robustness Physics-Guided Image Restoration
National Key Lab of Multispectral Information Intelligent Processing Technology School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China · School of Ocean Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China
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
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

Aug 10, 2026J/K move · Enter open · S save

Degraded Infrared Small Object Detection via Degradation-Adapted Physics-Guided Restoration

Xinkai Lu, Wenjun Chen, Yi Li, Yi Chang+1

National Key Lab of Multispectral Information Intelligent Processing Technology School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China · School of Ocean Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China

Image Corruption Robustness Physics-Guided Image Restoration Infrared Small Target Detection Image Restoration

Open paper AskPDF arXiv

Infrared small object detection has made significant progress in recent years. However, degradations such as fog and nonuniformity can suppress target-background contrast, substantially increasing detection difficulty. Existing methods mainly rely on image restoration as preprocessing, but they are typically designed for specific degradation types and fail to generalize to varying degradations. To alleviate this, we propose DAISOD, a degradation-adapted infrared small object detection framework for robust detection under different degradations. DAISOD first identifies the type and severity of degradations, then adapts the processing via dedicated branches, and finally fuses the results for subsequent detection. Moreover, a physics-guided restoration mechanism is incorporated to explicitly estimate degradation parameters and remove degradation effects through physical models, avoiding excessive restoration that may erase small targets. Moreover, we construct a degraded infrared small object detection dataset covering diverse degradation types and levels. Extensive experiments show that DAISOD outperforms state-of-the-art methods under various degradation conditions.

Abstract

Explore similar work

Degraded Infrared Small Object Detection via Degradation-Adapted Physics-Guided Restoration

Selective Attention-Based Network for Robust Infrared Small Target Detection

SANet: Selective Attention Network for Infrared Small Target Detection