Research map / Label-efficient and robust learning / Self-supervised and contrastive

Universal Image Restoration Pre-training via Degradation Classification

ICLR 2025 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Degradation Classification Pre-Training (DCPT), which pre-trains restoration encoders by classifying the degradation type of inputs from encoder features, then fine-tunes them for all-in-one, single-task, and mixed-degradation restoration.

Problem

Universal image restoration lacks effective pre-training; existing self-supervised methods (contrastive learning, masked image modeling) are not tailored to restoration and discard the decoder after pre-training.

Approach

Uses the degradation type of each restoration image, which is free supervision, as the pre-training target: a lightweight ResNet18 decoder classifies degradation from multi-level encoder features, alternated with a pixel restoration loss, and the decoder is kept for DC-guided cross-task transfer.

Key results

• DCPT-NAFNet reaches 29.84 dB average on 5D all-in-one vs 27.76 for NAFNet, and 29.72 dB on 10D all-in-one (+2.55 dB over NAFNet's 27.17).
• On CDD mixed degradation, DCPT-NAFNet improves NAFNet by 5.20 dB (low-light+haze+rain, 26.23 vs 21.03) and 5.58 dB (low-light+haze+snow).
• In zero-shot Gaussian denoising at sigma=50 on Urban100, DCPT-SwinIR improves SwinIR by 1.11 dB (27.63 vs 26.52).

Measured effects reported in the paper

Surprising findings

• Randomly initialized restoration networks already classify degradation type at 52-71% accuracy, rising to over 94% after all-in-one training.
• Degradation-classification-only pre-training (S_DC alone) still gives most of the gain, suggesting the benefit comes from the classification objective rather than generative ability.

Limitations

• Degradation classes are limited to single degradation types; mixed degradations are not directly addressed during classification.
• Evaluated mainly on synthetic degradations; real-world scenarios are left to future work.
• Decoder architecture choice matters and is not optimized for each restoration backbone.

Experimental setup

Datasets: Rain200L RESIDE BSD400 WED GoPro LOL RainH&RainL Snow100K DPDD CDD Urban100 Kodak24

Benchmarks: 5D all-in-one restoration 10D all-in-one restoration CDD11 mixed degradation Urban100 Kodak24 BSD68 (Gaussian denoising) UIEB GoPro HIDE Rain100L SOTS Test100L

Metrics: PSNR SSIM classification accuracy

Baselines: AirNet IDR InstructIR DACLIP DiffUIR-L PromptIR OneRestore DegAE TransWeather WeatherDiff

Builds on: Restormer PromptIR NAFNet masked image modeling Focal Loss

Closest papers by meaning

• DDR: Exploiting Deep Degradation Response as Flexible Image Descriptor (NeurIPS 2024 · similarity 0.86)
• Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.86)
• Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.85)
• Controlling Vision-Language Models for Multi-Task Image Restoration (ICLR 2024 · similarity 0.84)
• Learning Degradation-Independent Representations for Camera ISP Pipelines (CVPR 2024 · similarity 0.84)
• Vision-Language Gradient Descent-driven All-in-One Deep Unfolding Networks (CVPR 2025 · similarity 0.84)