Research map / Image generation / Super-resolution
Navigating Beyond Dropout: An Intriguing Solution towards Generalizable Image Super Resolution
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 a feature-statistics alignment regularizer (linear and RFF-based nonlinear) that aligns mean and covariance of features from same-content, different-degradation image pairs, as an alternative to Dropout for generalizable blind super-resolution.
Problem
Blind image super-resolution models overfit to training degradations and generalize poorly to unknown real-world degradations; Dropout, a prior regularizer, improves generalization but harms high-frequency detail.
Approach
Treat images with different degradations but identical content as different "styles" and align their first- and second-order feature statistics (mean and covariance), linearly or via random Fourier feature nonlinear mapping, as a model-agnostic regularizer that is complementary to degradation-pool data synthesis.
Key results
- Ours improves SRResNet by an average 0.57 dB vs 0.26 dB for Dropout, and RRDB by 0.72 dB vs 0.31 dB across six benchmarks.
- On NTIRE 2018 realistic data, Ours outperforms Dropout in PSNR for SRResNet and RRDB (e.g., RRDB mild 17.88 vs 17.31 dB).
- LPIPS improves with Ours while Dropout degrades LPIPS on several datasets (e.g., SRResNet Urban100 0.319 vs 0.333 for Dropout).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feature-statistics alignment to SRResNet instead of using no regularization | SRResNet · NTIRE 2018 SR challenge, mild | PSNR: 16.94 -> 17.69 | Helped |
| Add feature-statistics alignment to SRResNet instead of using no regularization | SRResNet · NTIRE 2018 SR challenge, difficult | PSNR: 17.84 -> 18.30 | Helped |
| Add feature-statistics alignment to SRResNet instead of using no regularization | SRResNet · NTIRE 2018 SR challenge, wild | PSNR: 17.55 -> 17.94 | Helped |
| Add channel-wise Dropout | SRResNet · NTIRE 2018 SR challenge, mild | PSNR: 16.94 -> 17.10 | Helped |
| Add channel-wise Dropout | SRResNet · Set5 | LPIPS: 0.241 -> 0.245 | Hurt |
| Replace channel-wise Dropout with feature-statistics alignment | SRResNet · NTIRE 2018 SR challenge, mild | PSNR: 17.10 -> 17.69 | Helped |
| Replace channel-wise Dropout with feature-statistics alignment | SRResNet · Set5 | LPIPS: 0.245 -> 0.238 | Helped |
| Add feature-statistics alignment to RRDB instead of using no regularization | RRDB · NTIRE 2018 SR challenge, mild | PSNR: 16.79 -> 17.88 | Helped |
| Add feature-statistics alignment to RRDB instead of using no regularization | RRDB · NTIRE 2018 SR challenge, difficult | PSNR: 17.66 -> 18.47 | Helped |
| Add feature-statistics alignment to RRDB instead of using no regularization | RRDB · NTIRE 2018 SR challenge, wild | PSNR: 17.38 -> 18.21 | Helped |
| Add channel-wise Dropout | RRDB · Set5 | LPIPS: 0.218 -> 0.229 | Hurt |
| Replace channel-wise Dropout with feature-statistics alignment | RRDB · NTIRE 2018 SR challenge, mild | PSNR: 17.31 -> 17.88 | Helped |
Surprising findings
- Dropout improves PSNR for blind SR but worsens LPIPS, indicating loss of high-frequency detail.
- Gains are largest for SRResNet and RRDB and small for MSRN and SwinIR (about 0.22 dB average).
Limitations
- Linear alignment and nonlinear RFF alignment rely on the authors' choice of mean and covariance statistics, which the paper admits lack solid theoretical foundation.
- Improvements are small for transformer and multi-scale baselines (MSRN, SwinIR).
- Requires paired images with same content but different degradations, which the degradation generator must produce.
Experimental setup
Datasets: DIV2K Set5 Set14 BSD100 Test2k Urban100 Manga109 NTIRE 2018 realistic data
Benchmarks: Set5 Set14 BSD100 Test2k Urban100 Manga109 NTIRE 2018
Metrics: PSNR LPIPS MAPE in frequency domain
Baselines: Dropout SRResNet RRDB MSRN SwinIR
Builds on: Dropout Kong et al. dropout for blind SR Deep CORAL style alignment Random Fourier features Real-ESRGAN degradation pipeline
Closest papers by meaning
- Reflash Dropout in Image Super-Resolution (CVPR 2022 · similarity 0.91)
- Adaptive Dropout: Unleashing Dropout across Layers for Generalizable Image Super-Resolution (CVPR 2025 · similarity 0.89)
- Learning the Non-Differentiable Optimization for Blind Super-Resolution (CVPR 2021 · similarity 0.88)
- GSDD: Generative Space Dataset Distillation for Image Super-resolution (AAAI 2024 · similarity 0.86)
- DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (CVPR 2025 · similarity 0.86)
- Noise-Free Optimization in Early Training Steps for Image Super-resolution (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.