EORestore-Agent: Fidelity-Guided Agentic Restoration of Remote Sensing Images with Composite Degradations

Heli Qi; Zeqi Zhou; Jingjun Yi; Kunyi Liu; Ziyang Lihe; Junjue Wang; Osamu Yoshie; Naoto Yokoya

cs.CVOct 5, 2026

EORestore-Agent: Fidelity-Guided Agentic Restoration of Remote Sensing Images with Composite Degradations

Authors: Heli Qi, Zeqi Zhou, Jingjun Yi, Kunyi Liu, Ziyang Lihe, Junjue Wang, Osamu Yoshie, Naoto Yokoya

Organizations: Graduate School of Information, Production and Systems, Waseda University, Kitakyushu, Fukuoka 808-0135, Japan · RIKEN Center for Advanced Intelligence Project (AIP), RIKEN, Tokyo 103-0027, Japan · Graduate School of Frontier Sciences, The University of Tokyo, Kashiwa, Chiba 277-8561, Japan · Informatics Institute, University of Amsterdam, Amsterdam 1098 XH, The Netherlands · Wuhan University, Wuhan 430072, China

Abstract

Remote sensing images often carry composite degradations, in which haze, cloud, noise, blur, low light, and low resolution coexist. Restoring them requires deciding which tool to apply, in what order, and when to stop, yet no clean reference is available at inference time to verify these decisions. All-in-one models trained on single degradations converge to a narrow PSNR band as degradations accumulate. To formulate real-world remote sensing restoration as a traceable trajectory, we present EORestore-Agent, which replaces this unmeasurable objective with reference-free, verifiable per-step decisions. A fine-tuned vision-language model reports all residual degradation types, whose tool pools are scored together, so the restoration order emerges from step-wise selection. A relative quality scorer, trained with full-reference supervision on synthetic degradation chains, predicts the changes in PSNR, SSIM, and LPIPS from the current image to each candidate. A step is accepted only when no predicted change is negative and the predicted PSNR gain is positive. Otherwise, the agent keeps the current image. On a synthetic Landsat-8 benchmark with six degradation types, EORestore-Agent improves PSNR by 2.3 to 3.2 dB over the strongest retrained all-in-one baseline on composites of two to six degradations, whereas zero-shot natural-image agents fall below the degraded input in PSNR in 17 of 18 settings. Replacing the learned scorer with no-reference quality differences costs 1.1 to 4.6 dB. The remaining harmful steps are small and cluster near the acceptance threshold. Sentinel-2 examples illustrate transfer to real atmospheric degradation without retraining.

Figures & tables

Fig. 1: Overview of EORestore-Agent. Top: the iterative restoration loop with perception, merged candidate pool, relative quality scoring, and three-metric gating. Middle: an example trajectory on a tile with four degradation types, showing the selected tool and PSNR at each step. Bottom: detailed view of the perception layer, execution layer, and toolbox.

Type	Tools	Size
Cloud	SpA-GAN [ 28 ] , ACA-CRNet [ 30 ] , MSDA-CR [ 29 , 31 ] , TGPNet [ 12 ]	4
Haze	DehazeFormer [ 32 ] , PSMB-Net [ 33 ] , RSDformer [ 34 ] , AU-Net [ 35 ] , Ada4DIR ${}_{\text{h}}$ , Restormer ${}_{\text{h}}$	6
Noise	Ada4DIR ${}_{\text{n}}$ , Restormer ${}_{\text{n}}$	2
Blur	Ada4DIR ${}_{\text{b}}$ , Restormer ${}_{\text{b}}$	2
Low light	Ada4DIR ${}_{\text{l}}$ , Restormer ${}_{\text{l}}$	2
Low resolution	FunSR [ 36 ]	1

TABLE I: Restoration Toolbox

Natural-image all-in-one
	Number of composited degradations
	1			2			3			4			5			6
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Degraded input	22.93	0.706	0.377	18.84	0.562	0.586	16.09	0.476	0.770	14.24	0.358	0.874	14.66	0.362	0.915	14.45	0.412	0.877
PromptIR [ 7 ]	31.82	0.887	0.193	19.40	0.709	0.411	17.29	0.597	0.597	16.29	0.537	0.696	15.82	0.482	0.774	15.61	0.478	0.809
Restormer [ 52 ]	32.64	0.889	0.196	22.81	0.744	0.392	19.04	0.629	0.585	18.13	0.574	0.661	17.12	0.530	0.730	16.35	0.513	0.775

TABLE II: Comparison on the synthetic benchmark. PSNR (dB,

\uparrow

), SSIM (

\uparrow

), and LPIPS (

\downarrow

) averaged over all degradation combinations at each

N

. Best and second best among restoration methods are in bold and underlined. The degraded input is not ranked. Ties share the mark.

Fig. 2: PSNR versus the number of composited degradations

N

, plotted from Table II . Each band spans the minimum and maximum of a baseline group, and the line with markers is the group median. The legend gives the number of models in each group.

Fig. 3: Qualitative comparison on the synthetic benchmark, one case per $N$ from 2 to 6. Row labels give the degradation combination. The PSNR of each output is shown at the bottom left, with the best value of each row in bold.

	$N{=}2$			$N{=}3$			$N{=}4$
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Restormer	24.91	0.796	0.272	22.23	0.686	0.488	20.44	0.549	0.670
Ada4DIR	21.03	0.760	0.312	20.00	0.662	0.492	19.06	0.551	0.679
PhyDAE	19.03	0.746	0.254	19.93	0.687	0.419	19.10	0.593	0.592
CoRE-UIR	20.55	0.748	0.267	20.04	0.673	0.438	19.45	0.578	0.593
Ours	26.27	0.834	0.255	22.77	0.698	0.442	21.07	0.572	0.626

TABLE III: Comparison on combinations composed entirely of seen degradation types. Only the four types covered by AIO training (haze, noise, blur, low light) are included. Restormer is the strongest baseline at

N\geq 2

in Table II . The

N{=}4

column contains a single combination. Best in bold, second best underlined. Ties share the mark.

			$N{=}1$			$N{=}2$			$N{=}3$
Perception	Selection signal	Fallback	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Single-label	$\hat{\Delta}_{P,S,L}$	$\checkmark$	33.58	0.881	0.184	25.90	0.769	0.365	21.89	0.646	0.546
Multi-label	$\hat{\Delta}_{P}$	$\checkmark$	33.66	0.874	0.194	26.05	0.762	0.376	22.06	0.640	0.556
Multi-label	$\hat{\Delta}_{P,S}$	$\checkmark$	33.67	0.880	0.194	26.05	0.770	0.370	22.07	0.647	0.551
Multi-label	$\hat{\Delta}_{P}$	$\times$	33.50	0.864	0.210	25.96	0.754	0.382	21.97	0.629	0.564
Multi-label	NIQE-diff	$\checkmark$	29.42	0.844	0.215	23.42	0.730	0.369	18.94	0.596	0.556

TABLE IV: Ablation on the Full Benchmark (270 Tiles). PSNR (dB,

\uparrow

), SSIM (

\uparrow

), and LPIPS (

\downarrow

) are averaged over all combinations at each

N

. Selection signal:

\hat{\Delta}_{P,S,L}

gates on all three predicted changes,

\hat{\Delta}_{P}

and

\hat{\Delta}_{P,S}

drop scorer heads, and X-diff ranks candidates by the change in the no-reference score X. Fallback

\times

removes the input image from the candidate set. Best in bold, second best underlined; ties share the mark.

Fig. 4: Real-world cases on Sentinel-2 L2A tiles. The reference is the L2A acquisition of the same area from a nearby cloud-free date, for visual comparison only (no full-reference metric). Row labels list the degradation set perceived at step 1. The red box marks the region enlarged at the bottom right of each cell.

Fig. 5: Trajectory statistics. (a) Mean true PSNR gain at each step, one curve per

N

. Terminated trajectories hold their final value. (b) Share of accepted steps with two or more active types. (c) Share of trajectories that end by abstention, by exhausting the repair budget, or because no degradation is perceived.

Fig. 6: Restoration trajectories, one per $N$ from 2 to 6. Rows 2 to 4 use the tiles of Fig. 3 , rows 1 and 5 a different tile of the same combination. Each solid arrow is an accepted step, labeled with its true PSNR change in dB. Under each frame, the strip marks the perceived types and the type of the applied tool, followed by that type, the tool, and its predicted PSNR change. Each trajectory ends with an abstention (dashed box). No tool is predicted to raise PSNR without worsening SSIM or LPIPS, so the agent keeps the current image and stops. Below the box is the declined tool, the one with the highest predicted PSNR gain. The three-metric gate rejected it because its predicted SSIM or LPIPS change was unfavorable.

$N$	Random	Ranking	Ours	Random	Ranking	Ours	Oracle	% of oracle
	Harmful rate (%) $\downarrow$			Gain (dB) $\uparrow$
1	46.7	40.0	27.9	3.97	7.68	7.69	8.28	93
2	43.2	30.4	24.7	1.04	3.63	3.57	4.40	81
3	48.9	35.5	27.1	0.07	1.87	1.84	2.57	72
4	48.2	33.8	26.6	0.29	1.65	1.72	2.32	74
5	41.4	28.2	21.4	0.29	1.30	1.22	2.16	56

TABLE V: Single-step comparison of selection rules on the same candidate pools. Harmful rate: share of steps whose chosen candidate lowers true PSNR. Gain: mean true PSNR change of the chosen candidate, over the steps where the full rule accepts a tool. Oracle picks the candidate with the highest true gain. The last column is the gain of Ours as a share of the oracle gain.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Natural-image all-in-one
	Bl			LL			Hz			Ns			Cl			LR
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Seen	$\checkmark$			$\checkmark$			$\checkmark$			$\checkmark$
Degraded input	34.10	0.812	0.402	15.79	0.590	0.233	14.18	0.799	0.249	25.19	0.458	0.565	17.11	0.816	0.318	31.23	0.759	0.496
PromptIR	36.43	0.904	0.171	40.93	0.995	0.005	36.42	0.982	0.025	35.91	0.895	0.154	17.24	0.815	0.305	23.97	0.734	0.499
Restormer	36.24	0.892	0.211	41.91	0.995	0.004	37.07	0.983	0.024	35.74	0.895	0.138	17.10	0.819	0.305	27.78	0.748	0.492

Appendix

TABLE VI: Per-combination results,

N{=}1

. Column labels: Cl cloud, Hz haze, Ns noise, Bl blur, LL low light, LR low resolution. A check in the Seen row marks combinations composed entirely of degradation types covered by AIO training. Method references are given in Table II .

Natural-image all-in-one
	Bl+LL			LL+Hz			Hz+Ns			Bl+LR			Cl+Ns			Cl+LR
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Seen	$\checkmark$			$\checkmark$			$\checkmark$
Degraded input	15.94	0.487	0.535	18.45	0.768	0.386	13.96	0.360	0.745	32.01	0.749	0.517	16.04	0.385	0.695	16.63	0.624	0.640
PromptIR	19.70	0.687	0.296	20.59	0.792	0.212	15.38	0.696	0.403	28.22	0.747	0.516	16.95	0.719	0.401	15.56	0.613	0.640
Restormer	26.08	0.790	0.306	26.02	0.871	0.189	22.62	0.727	0.322	29.87	0.753	0.512	16.87	0.716	0.388	15.38	0.609	0.635

Appendix

TABLE VII: Per-combination results,

N{=}2

Natural-image all-in-one
	Bl+LL+Hz			LL+Hz+Ns			Bl+Cl+LL			Bl+Cl+LR			Cl+Ns+LR			Hz+Ns+LR
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Seen	$\checkmark$			$\checkmark$
Degraded input	19.03	0.660	0.584	17.17	0.308	0.954	13.82	0.394	0.683	16.50	0.604	0.673	16.18	0.460	0.857	13.85	0.427	0.868
PromptIR	20.25	0.689	0.419	19.53	0.632	0.544	14.72	0.507	0.549	16.05	0.600	0.670	16.45	0.576	0.689	16.76	0.580	0.712
Restormer	22.63	0.749	0.402	21.82	0.622	0.573	16.21	0.625	0.519	15.76	0.599	0.670	15.78	0.569	0.684	22.05	0.608	0.660

Appendix

TABLE VIII: Per-combination results,

N{=}3

Natural-image all-in-one
	Bl+LL+Hz+Ns			Bl+Cl+LL+LR			Bl+Cl+Hz+LR			Bl+LL+Ns+LR			Cl+LL+Hz+Ns			Cl+Hz+Ns+LR
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Seen	$\checkmark$
Degraded input	17.33	0.254	1.033	14.21	0.389	0.741	11.94	0.514	0.758	15.51	0.313	0.874	14.52	0.295	0.937	11.95	0.385	0.904
PromptIR	19.29	0.576	0.663	16.39	0.588	0.676	17.35	0.595	0.682	15.65	0.391	0.738	15.59	0.563	0.644	13.47	0.512	0.774
Restormer	20.44	0.549	0.670	16.22	0.596	0.674	16.72	0.588	0.683	22.79	0.564	0.695	17.07	0.600	0.542	15.55	0.545	0.703

Appendix

TABLE IX: Per-combination results,

N{=}4

Natural-image all-in-one
	Bl+Cl+LL+Hz+Ns			Bl+Cl+LL+Hz+LR			Bl+Cl+LL+Ns+LR			Bl+Cl+Hz+Ns+LR			Bl+LL+Hz+Ns+LR			Cl+LL+Hz+Ns+LR
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Seen
Degraded input	14.74	0.226	1.030	14.91	0.529	0.791	14.09	0.282	0.897	11.79	0.355	0.917	17.48	0.401	0.922	14.92	0.381	0.931
PromptIR	15.93	0.518	0.725	16.07	0.504	0.736	14.04	0.342	0.806	13.45	0.491	0.799	19.45	0.533	0.781	16.00	0.501	0.798
Restormer	17.27	0.543	0.632	17.02	0.554	0.719	16.24	0.509	0.754	15.20	0.520	0.739	20.10	0.530	0.787	16.90	0.525	0.749

Appendix

TABLE X: Per-combination results,

N{=}5

Natural-image all-in-one
	Order 1			Order 2			Order 3			Order 4
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Seen
Degraded input	14.37	0.362	0.938	14.97	0.519	0.818	14.14	0.509	0.840	14.32	0.257	0.912
PromptIR	15.75	0.484	0.821	16.08	0.494	0.765	15.45	0.474	0.783	15.18	0.461	0.868
Restormer	16.66	0.508	0.774	16.89	0.542	0.736	16.11	0.529	0.754	15.76	0.472	0.838

Appendix

TABLE XI: Per-combination results,

N{=}6

. The four columns apply all six types in different orders: Order 1: Bl

\rightarrow

\rightarrow

\rightarrow

\rightarrow

\rightarrow

LR, Order 2: Ns

\rightarrow

\rightarrow

\rightarrow

\rightarrow

\rightarrow

Hz, Order 3: Ns

\rightarrow

\rightarrow

\rightarrow

\rightarrow

\rightarrow

Hz, Order 4: LR

\rightarrow

\rightarrow

\rightarrow

\rightarrow

\rightarrow

Ns. Column labels: Cl cloud, Hz haze, Ns noise, Bl blur, LL low light, LR low resolution. Method references are given in Table II .

Explore similar work

May 21, 2026cs.CV

OPERA: An Agent for Image Restoration with End-to-End Joint Planning-Execution Optimization

Real-world image restoration is challenging due to complex and interacting mixed degradations. Recent agent-based approaches address this problem by composing multiple task-specific restoration tools. However, empirical analysis reveals that their performance is fundamentally limited by implicitly constrained planning spaces and the lack of coordination among independently pretrained tools. To address these issues, we propose OPERA (Optimized Planning-Execution Restoration Agent), a framework that jointly optimizes restoration planning and tool execution in an end-to-end manner. On the planning side, OPERA uses reinforcement learning to directly optimize tool composition over a combinatorial plan space, with the final restoration quality as the reward. On the execution side, OPERA introduces agent-guided co-training of restoration tools, enabling them to learn cooperative behaviors under sequential composition. Extensive experiments on multi-degradation benchmarks and real-world datasets demonstrate that OPERA consistently outperforms both all-in-one restoration models and existing agent-based methods across diverse and complex degradation scenarios.

Feng Zhu, Shuyang Xie, Yihan Zeng +2

Harbin Institute of Technology · Huawei Noah’s Ark Lab

Image Restoration Restoration Iterative Co-Training

Ask PDF

Mar 29, 2026cs.CV

TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

Vision-language agents that orchestrate specialized tools for image restoration (IR) have emerged as a promising method, yet most existing frameworks operate in a training-free manner. They rely on heuristic task scheduling and exhaustive tool traversal, resulting in sub-optimal restoration paths and prohibitive computational cost. We argue that the core bottleneck lies in the absence of a learned policy to make decision, as a vision-language model cannot efficiently handle degradation-aware task ordering and tool composition. To this end, we propose TIR-Agent, a trainable image restoration agent that performs a direct tool-calling policy through a two-stage training pipeline of supervised fine-tuning (SFT) followed by reinforcement learning (RL). Two key designs underpin effective RL training: (i) a random perturbation strategy applied to the SFT data, which broadens the policy's exploration over task schedules and tool compositions, and (ii) a multi-dimensional adaptive reward mechanism that dynamically re-weights heterogeneous image quality metrics to mitigate reward hacking. To support high-throughput, asynchronous GPU-based tool invocation during training, we further develop a globally shared model-call pool. Experiments on both in-domain and out-of-domain degradations show that TIR-Agent outperforms 12 baselines, including 6 all-in-one models, 3 training-free agents, and 3 proprietary models, and achieves over 2.5

\times

inference speedup by eliminating redundant tool executions.

Guoli Jia, Yisheng Zhang, Haote Hu +11

Tsinghua University, Beijing, China · Data Science & Artificial Intelligence Research Institute, China Unicom, Beijing, China · Hunan University, Changsha, China +2

Image Restoration Offline Reinforcement Learning Cpu-Gpu Hybrid Designs

Ask PDF

May 29, 2026cs.CV

DiTTo: Scalable Order-aware All-in-One Image Restoration Agent

Real-world images rarely suffer from a single degradation, and the order in which degradations are removed substantially affects the final restoration quality, motivating agent-based image restoration (IR), where a vision-language model schedules a pool of pre-built restoration-experts. However, existing training-based agents require

\mathcal{O}((N^{\mathbf{D}})^{2})

restoration-expert calls per image to construct the Optimal Restoration-action Trajectory Dataset (ORTD), where

N^{\mathbf{D}}

denotes the number of degradation types in the universe

\mathbf{D}

, and couple agent training to a fixed restoration-expert pool, preventing extension to newly introduced restoration-experts without full retraining. To overcome these efficiency and extensibility bottlenecks, we propose \textbf{DiTTo}, a novel order-aware image restoration agent framework consisting of the DiTTo Simulator and the DiTTo Agent. The DiTTo Simulator combines

\cup

S-IR for single-step restoration-action simulation and AiO-IQA for per-action quality prediction, reducing ORTD construction to

\mathcal{O}(N^{\mathbf{D}})

simulator calls per image; the DiTTo Agent is trained by SFT on the simulator-generated ORTD, followed by \textbf{Order-aware Restoration Alignment (ORA)} that aligns degradation identification, restoration-action-ordering, and output format along independent axes. This enables \textbf{plug-and-play scalable extensibility}: adding a new restoration-expert requires updating only the lightweight ORA stage. On the MiO-100 evaluation set with up to five concurrent degradations, our DiTTo Agent achieves state-of-the-art multi-degradation restoration quality among previous agent-based IR methods.

Seungho Choi, Jihyong Oh

CMLab, Chung-Ang University

Image Restoration Restoration

Ask PDF