ClearAIR: A Human-Visual-Perception-Inspired All-in-One Image Restoration

Xu Zhang; Huan Zhang; Guoli Wang; Qian Zhang; Lefei Zhang

Source-linked AI summary

ClearAIR: A Human-Visual-Perception-Inspired All-in-One Image Restoration

Xu Zhang, Huan Zhang, Guoli Wang, Qian Zhang, Lefei Zhang

arXiv:2601.02763v2Jan 6, 2026cs.CV

TL;DR

Existing all-in-one restoration methods face challenges from spatially varying degradations and the need for separate or degradation-specific processing. ClearAIR addresses this with a Human Visual Perception-inspired coarse-to-fine framework combining global assessment, regional guidance, task recognition, and self-supervised internal clue reuse. The authors report state-of-the-art performance on synthetic and real-world datasets, while noting sensitivity to processing order and limited adaptability to non-uniform real-world degradations.

Problem
Existing restoration approaches can require separate models or use uniform processing despite spatially varying degradation and region-dependent restoration difficulty.
Method
ClearAIR uses an HVP-inspired hierarchical pipeline with MLLM-based IQA, semantic guidance, task identification, and self-supervised internal clue reuse for fine-detail reconstruction.
Results
ClearAIR achieves state-of-the-art performance on both synthetic and real-world datasets.
Takeaways & Limitations
The framework combines global quality assessment and local degradation understanding with internal image cues to support coarse-to-fine restoration.
Takeaways & Limitations
The fixed perception pipeline is sensitive to processing order and lacks adaptability to complex real-world degradations that vary spatially in type and intensity.

Abstract

from arXiv · show

All-in-One Image Restoration (AiOIR) has advanced significantly, offering promising solutions for complex real-world degradations. However, most existing approaches rely heavily on degradation-specific representations, often resulting in oversmoothing and artifacts. To address this, we propose ClearAIR, a novel AiOIR framework inspired by Human Visual Perception (HVP) and designed with a hierarchical, coarse-to-fine restoration strategy. First, leveraging the global priority of early HVP, we employ a Multimodal Large Language Model (MLLM)-based Image Quality Assessment (IQA) model for overall evaluation. Unlike conventional IQA, our method integrates cross-modal understanding to more accurately characterize complex, composite degradations. Building upon this overall assessment, we then introduce a region awareness and task recognition pipeline. A semantic cross-attention, leveraging semantic guidance unit, first produces coarse semantic prompts. Guided by this regional context, a degradation-aware module implicitly captures region-specific degradation characteristics, enabling more precise local restoration. Finally, to recover fine details, we propose an internal clue reuse mechanism. It operates in a self-supervised manner to mine and leverage the intrinsic information of the image itself, substantially enhancing detail restoration. Experimental results show that ClearAIR achieves superior performance across diverse synthetic and real-world datasets.

Introduction

ClearAIR addresses limitations in all-in-one restoration by using a Human Visual Perception-inspired, coarse-to-fine pipeline that combines global assessment, regional degradation analysis, and internal image cues for detail recovery.

Existing restoration methods often require separate models for each degradation type, increasing inference complexity and computational cost.
AiOIR methods improve unified restoration but commonly rely on degradation-specific representations and overlook spatially varying degradation.
ClearAIR progressively refines images from global structure to fine local details, following a Human Visual Perception-inspired coarse-to-fine strategy.
Its pipeline uses MLLM-based IQA for overall quality, semantic guidance and task identification for regional degradation analysis, and ICRM for fine-detail recovery.

Related Work

Related work motivates spatially aware, perception-inspired restoration because degradation and restoration difficulty vary across image regions. ClearAIR combines global understanding with local perception to preserve naturalness and semantic consistency under complex degradation.

All-in-one restoration seeks to handle diverse degradation types with a single unified model, while recent methods use prompts and large-scale vision models to improve complex restoration.
Most all-in-one methods process images uniformly, despite spatially variable degradation and differing restoration difficulty across flat and textured regions.
Human visual perception first interprets an image as a whole before analyzing its constituent parts, motivating hierarchical restoration designs.
ClearAIR integrates global coarse-grained understanding with local fine-grained perception to improve visual naturalness and maintain semantic consistency under complex degradation.
MLLM-based image quality assessment provides a reference-free quality score, aligning with practical restoration settings that lack high-quality reference images.

Methodology

ClearAIR uses a hierarchical restoration pipeline that combines overall quality assessment, region awareness, task recognition, and self-supervised internal clue reuse. These components guide restoration from global structure toward region-specific degradation handling and fine-detail recovery.

Framework overview: ClearAIR comprises MLLM-based IQA, semantic guidance, task identification, and internal clue reuse for coarse-to-fine restoration.The framework uses quality assessment, semantic masks, degradation prompts, and self-supervised internal cues to guide restoration.
Overall Assessment: An MLLM-based IQA model fuses visual and textual tokens, extracting quality reasoning to guide the restoration backbone through score embeddings.The image is encoded into visual tokens, compressed, fused with textual tokens, and processed by an MLLM; an intermediate state is transformed for the QGM.
Region Awareness: The Semantic Guidance Unit generates region masks from a pre-trained segmentation model and fuses semantic features through Semantic Cross-Attention.Mask Average Pooling integrates masked shallow features, while mask dropout improves robustness to varying mask quality before SCA-based guidance.
Task Recognition: The Task Identifier predicts local degradation types with DA-CLIP and converts degradation embeddings into prompts for the Degradation-Aware Module.Content and degradation embeddings support cross-attention for content-aware spatial enhancement, while degradation prompts modulate restoration features.
Internal Clue Reuse Mechanism: The Internal Clue Reuse Mechanism applies weak and strong augmentations to restored outputs and minimizes their L2 distance as an internal consistency signal.This self-supervised mechanism is designed to preserve fine details in restored images; its loss contribution is weighted by γ, initialized to 0.05.

Experiments

ClearAIR is evaluated under All-in-One and Single-task protocols across multiple restoration settings, with results showing strong performance across synthetic and real-world degradations. Ablations indicate that perception order and structured priors materially affect performance.

Experimental Setup: ClearAIR is evaluated under All-in-One and Single-task settings, with the All-in-One protocol covering four task configurations.The experiments use mixed-degradation training and task-specific testing for All-in-One evaluation.
Three Degradations: 31.08 dB PSNR on SOTS surpasses VLU-Net’s 30.71 dB in the Three Degradations task.ClearAIR also achieves the best average performance, with gains in high-noise removal and severe haze reduction.
Five Degradations: 30.45 dB average PSNR and 0.916 average SSIM establish the best averages in the Five Degradations task.ClearAIR performs particularly strongly in deblurring, while remaining competitive in low-light enhancement and denoising.
All-Weather and Composited Degradation: ClearAIR gains 0.22 dB over Histoformer on All-Weather and 0.62 dB over OneRestore on composited degradations.The corresponding evaluations report clearer outputs, artifact removal, and preservation of details and textures.
Ablation Studies: The proposed perception order achieves the best overall performance, whereas Where-What-How performs worst in the order ablation.The authors associate the poorer order with disrupting structural integrity by processing regional semantics first.
Ablation Studies: Removing structured MLLM-IQA, SGU, or task-identifier priors degrades performance, while removing LCRM also reduces performance to a lesser extent.These ablations support explicit prior modeling and the use of internal image structures for guidance and task adaptation.

Conclusion

ClearAIR concludes with a hierarchical HVP-inspired restoration framework that combines global assessment, regional understanding, task recognition, and internal clue reuse. Its experiments report state-of-the-art performance across synthetic and real-world datasets, while ablations evaluate the contributions of its components.

Framework: ClearAIR uses a coarse-to-fine HVP-inspired strategy that progresses from overall assessment to regional analysis and fine-detail restoration.The framework combines an MLLM-based IQA model, semantic guidance, task identification, and internal clue reuse.
Framework: The MLLM-based IQA, semantic guidance unit, and task identifier support degradation localization and understanding across regions.These components provide global quality assessment and local semantic and degradation cues.
Fine-Detail Restoration: The internal clue reuse mechanism uses self-supervised learning to improve recovery of detailed textures.Ablation evidence shows that removing LCRM reduces performance, though less than removing several structured priors.
Conclusion: Experimental results demonstrate state-of-the-art performance on synthetic and real-world datasets.The conclusion presents this outcome as the overall experimental result of ClearAIR.

A. More Details on Datasets and Evaluation

The evaluation organizes datasets into All-in-One mixed-degradation settings and Single-task settings, using both reference-based and no-reference image-quality metrics. The All-in-One benchmarks cover progressively broader combinations of noise, haze, rain, blur, low-light, snow, and raindrops.

All-in-One Setting: All-in-One training combines multiple degradation datasets, while testing is performed separately on datasets containing a single degradation type.The paper summarizes four All-in-One configurations.
All-in-One Setting: The Three Degradations setting includes Gaussian noise, haze, and rain, while Five Degradations adds motion blur and low-light enhancement.The listed datasets include BSD400, WED, BSD68, SOTS, Rain100L, GoPro, and LOL.
All-in-One Setting: The All-Weather setting covers haze and rain, raindrops, and snow, whereas Composited Degradations combines haze, rain, low-light, and snow on CDD-11.The composited benchmark synthesizes degradations on DIV2K.
Single-task Setting: Single-task experiments train and test one restoration task at a time using task-specific datasets for denoising, dehazing, and deraining.Examples include Gaussian noise levels σ ∈{15, 25, 50}, RESIDE-β for dehazing, and Rain100L for deraining.
Evaluation Metrics: Evaluation uses PSNR and SSIM as reference-based metrics and MUSIQ and NIQE as no-reference metrics.Higher PSNR, SSIM, and MUSIQ indicate better quality, while lower NIQE indicates better quality.

B. More Experiment Results

ClearAIR generalizes across real-world weather, denoising, dehazing, and deraining settings, achieving strong quantitative and qualitative results against competing methods.

Real-World All-Weather Task: ClearAIR achieves the best performance on both RainDS and Snow100K-real in the real-world All-Weather task.On Snow100K-real, evaluation uses MUSIQ and NIQE because ground-truth images are unavailable.
Single Degradation Task: ClearAIR outperforms ADFNet by 0.18/0.30 dB and FSNet by 0.34/0.96 dB in PSNR at noise level 15 on CBSD68 and Urban100.
Single Degradation Task: ClearAIR surpasses PromptIR by 0.60 dB in dehazing and 1.48 dB in deraining PSNR.It also improves over DehazeFormer by 0.13 dB and DRSformer by 0.38 dB.
Single Degradation Task: ClearAIR retains relatively more texture details in restored images, producing visually more pleasing results in dehazing and deraining comparisons.

C. Limitation and Future Work

ClearAIR remains limited by a fixed perception pipeline and component-level weaknesses in difficult real-world conditions; the paper proposes JND-aware adaptation as a future direction.

Limitations: ClearAIR’s fixed perception pipeline is inflexible when degradation types and intensities vary spatially in complex real-world scenarios.Its performance is sensitive to the processing order used in ablation studies.
Limitations: The semantic guidance module can produce inaccurate region masks under extreme low-visibility conditions.
Limitations: The internal detail enhancement uses a perceptually uninformed augmentation strategy that can yield suboptimal fine-detail recovery.
Future Work: The proposed future direction integrates Just Noticeable Difference mechanisms to adapt restoration routing and allocate computation to perceptually critical regions.The envisioned framework uses perceptual thresholds and JND-weighted mechanisms.