Beyond Model Design: Data-Centric Training and Self-Ensemble for Gaussian Color Image Denoising

1. Problem Definition

The paper addresses the NTIRE 2026 Image Denoising Challenge, specifically focusing on Gaussian color image denoising at a fixed noise level of $\sigma = 50$.

• Goal: To restore a clean image $y$ from a noisy observation $x$, where the degradation follows $x = y + n$ and $n \sim \mathcal{N}(0, \sigma^2I)$.
• Context: While deep learning has established mature architectures for this task (e.g., DnCNN, SwinIR, Restormer), the authors hypothesize that the remaining performance gap is not due to architectural limitations but rather to the statistical priors available in the training corpus and the efficiency of the inference strategy.

2. Methodology

The authors propose a solution that does not modify the underlying model architecture. Instead, they maximize the potential of the existing Restormer backbone through two complementary strategies: Data-Centric Training and Test-Time Enhancement.

A. Backbone Architecture

• Model: Restormer [63], a hierarchical encoder-decoder transformer using multi-DConv head transposed attention and gated convolutional feed-forward blocks.
• Strategy: The architecture is used exactly as published, without any structural changes. The focus is entirely on the training recipe and inference wrapper.

B. Data-Centric Training (Two-Stage Optimization)

The authors expand the standard training recipe significantly to inject stronger statistical priors into the model.

• Corpus Expansion: Instead of the standard 4-dataset mix (DIV2K, Flickr2K, BSD500, WED), the training pool is expanded to 7 sources containing 143,679 images:
  • Base: DIV2K, Flickr2K, OST, LSDIR.
  • Expanded: LIU4K-v2, NKUSR8K (from DiT4SR project), and DIV8K.
  • Preprocessing: Ultra-high-resolution images are cropped into ~2K sub-images to manage distribution while retaining rich texture statistics.
• Two-Stage Training Schedule:
  • Stage I: Fine-tuning on the base expanded corpus (DIV2K, Flickr2K, OST, LSDIR) to establish a stable restoration prior.
    • Hyperparameters: AdamW optimizer, MSE loss, LR $1\times10^{-4}$, Batch size 4 $\to$ 2 $\to$ 1.
  • Stage II: Further fine-tuning introducing the remaining diverse datasets (LIU4K-v2, NKUSR8K, DIV8K) to broaden generalization.
    • Hyperparameters: LR reduced to $1\times10^{-5}$, Batch size 4.
  • Initialization: Training starts from the public Restormer $\sigma=50$ pretrained weights, not random initialization.

C. Test-Time Enhancement

• Geometric Self-Ensemble ($\times8$): The model performs inference on the original image and 7 geometric transformations (flips and transposes) representing the dihedral group $D_4$. The final prediction is the average of the inverse-transformed outputs.
  • Formula: $\hat{y} = \frac{1}{8} \sum_{k=1}^{8} T_k^{-1}(f_\theta(T_k(x)))$.
  • Trade-off: Increases inference time by $\approx 8\times$ but improves stability and accuracy.
• TLC-Style Wrapper: A local inference wrapper (from TLC [21]) is retained for implementation consistency. However, ablation studies show its quantitative contribution is negligible in this specific setting.

3. Key Contributions

1. Performance without Architecture Changes: Demonstrates that a mature Restormer-based denoiser can be significantly improved (up to 3.366 dB PSNR gain) solely through corpus expansion and a two-stage training schedule, without altering the backbone.
2. Systematic Quantification of Gains: Provides a rigorous ablation study separating the contributions of:
   • Stronger training checkpoints (the dominant factor).
   • Geometric self-ensemble (small but consistent boost).
   • The TLC-style wrapper (negligible impact).
3. Competitive Validation: The method achieved 2nd place on the NTIRE 2026 leaderboard (Team: wedream), validating the effectiveness of the data-centric approach under competitive conditions.

4. Experimental Results

The evaluation was conducted on a unified 100-image validation set.

• Final Performance:
  • PSNR: 30.762 dB
  • SSIM: 0.861
  • Inference Time: ~8815 ms (with $\times8$ ensemble)
  • Peak Memory: ~36,956 MB
• Comparison with Baseline:
  • Compared to the public Restormer $\sigma=50$ pretrained baseline, the proposed method achieved a PSNR improvement of +3.366 dB (with ensemble) and +0.0737 SSIM.
• Ablation Insights:
  • Training vs. Inference: The vast majority of the gain comes from the stronger checkpoint derived from the expanded corpus and two-stage training.
  • Self-Ensemble: Contributes a small but stable gain of ~0.027 dB PSNR and 0.0004 SSIM.
  • TLC Wrapper: Shows no measurable quantitative contribution in this specific Gaussian denoising task.
  • Checkpoint Evolution: The model converges strongly before the final stage ends, with only a marginal gain (0.0042 dB) between the intermediate checkpoint (200k iterations) and the final model.

5. Significance and Conclusion

This paper challenges the prevailing trend of constantly designing new, complex restoration backbones. It argues that for mature tasks like Gaussian denoising at fixed noise levels:

• Data is King: Expanding the diversity and scale of the training corpus yields far greater returns than architectural tweaks.
• Optimization Matters: A structured, two-stage training schedule allows the model to better absorb the statistical priors of diverse datasets.
• Honest Engineering: By isolating the effects of inference wrappers and ensembles, the authors clarify that "engineering tricks" (like the TLC wrapper) often persist in pipelines without adding value, whereas robust training data is the true driver of performance.

The work serves as a strong benchmark for data-centric AI, proving that maximizing the utility of existing SOTA architectures through rigorous data engineering can still yield state-of-the-art results.