Compositional-Degradation UAV Image Restoration: Conditional Decoupled MoE Network and A Benchmark

Jinquan Yan; Zhicheng Zhao; Zhengzheng Tu; Chenglong Li; Jin Tang; Bin Luo

eess.IVApr 10, 2026

Compositional-Degradation UAV Image Restoration: Conditional Decoupled MoE Network and A Benchmark

Authors: Jinquan Yan, Zhicheng Zhao, Zhengzheng Tu, Chenglong Li, Jin Tang, Bin Luo

Organizations: Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, School of Computer Science and Technology, Anhui University, Hefei 230601, China. · Key Laboratory of Intelligent Computing & Signal Processing (Anhui University), Ministry of Education, Anhui Provincial Key Laboratory of Multimodal Cognitive Computation, School of Artificial Intelligence, Anhui University, Hefei 230601, China.

Abstract

UAV images are critical for applications such as large-area mapping, infrastructure inspection, and emergency response. However, in real-world flight environments, a single image is often affected by multiple degradation factors, including rain, haze, and noise, undermining downstream task performance. Current unified restoration approaches typically rely on implicit degradation representations that entangle multiple factors into a single condition, causing mutual interference among heterogeneous corrections. To this end, we propose DAME-Net, a Degradation-Aware Mixture-of-Experts Network that decouples explicit degradation perception from degradation-conditioned reconstruction for compositional UAV image restoration. Specifically, we design a Factor-wise Degradation Perception module(FDPM) to provide explicit per-factor degradation cues for the restoration stage through multi-label prediction with label-similarity-guided soft alignment, replacing implicit entangled conditions with interpretable and generalizable degradation descriptions. Moreover, we develop a Conditioned Decoupled MoE module(CDMM) that leverages these cues for stage-wise conditioning, spatial-frequency hybrid processing, and mask-constrained decoupled expert routing, enabling selective factor-specific correction while suppressing irrelevant interference. In addition, we construct the Multi-Degradation UAV Restoration benchmark (MDUR), the first large-scale UAV benchmark for compositional UAV image restoration, with 43 degradation configurations from single degradations to four-factor composites and standardized seen/unseen splits.Extensive experiments on MDUR demonstrate consistent improvements over representative unified restoration methods, with greater gains on unseen and higher-order composite degradations. Downstream experiments further validate benefits for UAV object detection.

Figures & tables

Fig. 1: Real UAV examples illustrating the impact of adverse degradations on downstream object detection. The examples include single-degradation cases and composite cases. Colored solid boxes denote detector predictions, and red dashed boxes highlight visually identifiable missed targets. Adverse degradations substantially reduce detection reliability, with more severe failures in composite cases.

Fig. 2: Motivation of DAME-Net. In blind unified restoration, multiple active degradations are often compressed into a single holistic condition, which blurs factor boundaries and limits factor-specific correction. DAME-Net instead performs explicit degradation perception and uses factor-wise cues for conditioned restoration.

Fig. 3: Motivation for FDPM. Conventional hard alignment forces a composite degradation (e.g., rain+haze) to match a single prompt, whereas FDPM preserves its overlap with constituent factors through label-similarity-guided soft alignment.

Fig. 4: Overview of DAME-Net. FDPM predicts a degradation mask $\hat{m}$ and an image embedding $p$ , which are transformed by the Degradation Encoder into stage-wise conditioning vectors $\{g_{s}\}_{s=1}^{5}$ . CDMM then performs degradation-conditioned restoration, while the Low Frequency Predictor is supervised by Guided Filter Loss and the residual path is supervised by Spatial $\ell_{1}$ Loss and Masked Frequency $\ell_{1}$ Loss.

Fig. 5: Architecture of CDCB. The frequency branch transforms input features via FFT, applies degradation-conditioned spectral modulation with a rank- $K$ frequency mask, and refines the DC component via content-adaptive offset prediction. The spatial branch applies Swin window attention. A learned gate $w$ balances the two branches.

Fig. 6: Architecture of the DC-MoE. The predicted degradation mask first activates degradation-relevant experts in a Global Group and a Spatial Group, while suppressing irrelevant experts. Routing weights are then assigned only over the activated expert subset. The Spatial Group further employs a pixel-wise Spatial Router to localize expert responses within the feature map.

Seen Settings
Setting	Ours		PromptIR [ 22 ]		DehazeFormer [ 20 ]		AirNet [ 24 ]		Restormer [ 19 ]		AdaIR [ 23 ]
Setting	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
Avg. Seen Single	29.52	0.9091	29.36	0.9066	28.93	0.9002	27.40	0.8347	29.16	0.9041	28.40	0.8931
Avg. Seen Double	26.88	0.8389	26.61	0.8328	26.22	0.8207	25.07	0.7454	26.43	0.8283	25.79	0.8102
Avg. Seen Triple	25.73	0.8105	25.39	0.7988	24.95	0.7804	24.36	0.7048	25.25	0.7927	24.65	0.7681
Overall Seen	27.67	0.8602	27.43	0.8544	27.01	0.8433	25.82	0.7717	27.25	0.8504	26.57	0.8338

TABLE I: Quantitative comparison on MDUR using averaged results only (PSNR/SSIM). Results are reported for in-distribution (seen) and zero-shot (unseen) degradation settings. Best and second-best results are marked in bold and underlined , respectively.

Method	Precision $\uparrow$	Recall $\uparrow$	mAP50 $\uparrow$	mAP50-95 $\uparrow$	$\Delta$ mAP50 $\uparrow$
GT	0.5851	0.5486	0.5419	0.3211	–
Degraded	0.1530	0.1046	0.0971	0.0542	–
AdaIR	0.3189	0.2343	0.2216	0.1259	0.1245
DehazeFormer	0.3269	0.2392	0.2289	0.1307	0.1318
Restormer	0.3434	0.2505	0.2420	0.1380	0.1449
PromptIR	0.3495	0.2547	0.2469	0.1412	0.1497

TABLE II: Downstream object detection on all 43 degraded UAV settings. “Degraded” denotes detection on degraded inputs, “GT” the upper bound on clean images, and

\Delta

mAP50 the improvement over Degraded. Best and second-best results are marked in bold and underlined , respectively.

Fig. 7: Qualitative comparison on MDUR under single and composite degradations. From left to right: Input, Restormer, PromptIR, Ours, and Ground Truth (GT). Our method better preserves scene structure and color consistency while leaving fewer visible residual degradations.

Variant	All		Quad
Variant	PSNR	SSIM	PSNR	SSIM
Full model (Ours)	23.04	0.7410	16.98	0.5211
GT Mask (oracle mask)	23.04	0.7415	16.97	0.5226
w/o Semantic Embedding	21.58	0.7055	13.67	0.4356
w/o Global Token	21.96	0.7201	14.41	0.4707
w/o Strict Token Masking	22.48	0.7320	15.96	0.5060

TABLE III: Ablation of degradation representation components. “GT Mask” replaces the predicted binary mask with the ground-truth mask while keeping the CLIP-derived semantic embedding unchanged. “All” and “Quad” denote averages over all 43 tasks and the quad-degradation subset, respectively. Best and second-best results are marked in bold and underlined , respectively.

Variant	All		Quad
Variant	PSNR	SSIM	PSNR	SSIM
Full model (Ours)	23.04	0.7410	16.98	0.5211
w/o Freq-Spa Gate	22.09	0.7203	15.22	0.4765
w/o DC-MoE	22.34	0.7222	15.48	0.4804
w/o Decouple Gate	22.46	0.7284	15.55	0.4836
w/o Freq Branch	22.28	0.7278	16.12	0.5108

TABLE IV: Ablation of restoration architecture components. “All” and “Quad” denote averages over all 43 tasks and the quad-degradation subset, respectively. Best and second-best results are marked in bold and underlined , respectively.

Variant	All		Quad
Variant	PSNR	SSIM	PSNR	SSIM
Full model (Ours)	23.04	0.7410	16.98	0.5211
w/o Freq Loss	22.32	0.7304	14.49	0.4790
w/o Mask Overload	22.58	0.7287	16.54	0.4910
w/o Guided Filter Loss	22.91	0.7452	16.88	0.5337

TABLE V: Ablation of training strategies. “All” and “Quad” denote averages over all 43 tasks and the quad-degradation subset, respectively. Best and second-best results are marked in bold and underlined , respectively.

Explore similar work

CoRE-UIR: Prior-guided common and residual experts for efficient all-in-one remote sensing image restoration
Jul 30, 2026Zaiyan Zhang, Qiangqiang Yuan, Jie Li +5Mixture of Experts All-in-One Image Restoration
School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, Hubei, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, Hubei, China · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, Hubei, China
DVANet: Degradation-aware Visual-prior Alignment Network for Image Restoration
Jun 17, 2026Yanjie Tu, Qingsen Yan, Axi Niu +3All-in-One Image Restoration Image Restoration
School of Computer Science, Northwestern Polytechnical University, Xi’an 710000, China · Shenzhen Research Institute of Northwestern Polytechnical University, Shenzhen 518057, China · State Key Laboratory of Internet of Things for Smart City, University of Macau, Macau 999078, China
Degradation-Aware Adaptive Context Gating for Unified Image Restoration
May 2, 2026Lei He, Jielei Chu, Fengmao Lv +4All-in-One Image Restoration Image Restoration
the School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China · the Engineering Research Center of Sustainable Urban Intelligent Transportation, Ministry of Education, Chengdu, Sichuan 611756, China · School of Computing and Artificial Intelligence, Jiangxi University of Finance and Economics, Nanchang, 330013, China +3

Jul 30, 2026J/K move · Enter open · S save

CoRE-UIR: Prior-guided common and residual experts for efficient all-in-one remote sensing image restoration

Zaiyan Zhang, Qiangqiang Yuan, Jie Li, Ziyang Lihe+4

School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, Hubei, China · School of Artificial Intelligence, Wuhan University, Wuhan 430072, Hubei, China · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, Hubei, China

Mixture of Experts All-in-One Image Restoration Image Restoration

Open paper AskPDF arXiv

Remote sensing images acquired by unmanned aerial vehicles (UAVs) and satellites are often degraded by adverse weather, illumination variation, and imaging artifacts, which may co-occur and jointly induce global distribution shifts and local structural corruption. Although All-in-One image restoration offers an appealing unified alternative to task-specific pipelines, existing methods still suffer from weak or implicit degradation cues and parameter redundancy caused by full-rank multi-expert designs with overlapping restoration behaviors. We propose CoRE-UIR (Common and Residual Experts for Universal Image Restoration), a prior-guided global-local framework centered on the Common-and-Residual Expert Block (CoRE). CoRE explicitly decomposes restoration capacity into a common dense expert for degradation-invariant restoration and low-rank residual experts for degradation-specific compensation, enabling adaptive specialization without redundant expert replication. Built on this design, Degradation Prior Embedding (DPE) adapts frozen CLIP features into an explicit restoration-oriented prior, while Global Feature Modulation (GFM) aligns global feature statistics before local residual compensation. We also construct MDVD-108K (Multi-Degradation VisDrone), a large-scale UAV restoration dataset covering both single and compound degradations, together with a real-world test set. Extensive experiments on multiple datasets show that CoRE-UIR improves the overall average PSNR by 1.05 dB while running 11.83

\times

faster and reducing peak memory by 85.3% relative to the strongest baseline, BaryIR, thereby maintaining a favorable quality-efficiency trade-off. Evaluations on downstream tasks and unseen degradation also validate the generalizability of CoRE-UIR. The code and dataset will be released at https://github.com/zzaiyan/CoRE-UIR.

Abstract

Explore similar work

CoRE-UIR: Prior-guided common and residual experts for efficient all-in-one remote sensing image restoration

DVANet: Degradation-aware Visual-prior Alignment Network for Image Restoration

Degradation-Aware Adaptive Context Gating for Unified Image Restoration