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

TL;DR

This paper enhances Gaussian color image denoising by improving data-centric training and test-time self-ensemble techniques on the Restormer architecture, achieving significant performance gains.

Contribution

It introduces a two-stage training process with larger datasets and applies geometric self-ensemble at inference, boosting denoising performance without new model design.

Findings

Expanded training data significantly improves denoising performance.

Two-stage optimization yields up to 3.366 dB PSNR improvement.

Self-ensemble provides marginal but consistent gains.

Abstract

This paper presents our solution to the NTIRE 2026 Image Denoising Challenge (Gaussian color image denoising at fixed noise level $\sigma = 50$). Rather than proposing a new restoration backbone, we revisit the performance boundary of the mature Restormer architecture from two complementary directions: stronger data-centric training and more complete Test-Time capability release. Starting from the public Restormer $\sigma\!=\!50$ baseline, we expand the standard multi-dataset training recipe with larger and more diverse public image corpora and organize optimization into two stages. At inference, we apply $\times 8$ geometric self-ensemble to further release model capacity. A TLC-style local inference wrapper is retained for implementation consistency; however, systematic ablation reveals its quantitative contribution to be negligible in this setting. On the challenge validation set of 100 images, our final submission achieves 30.762 dB PSNR and 0.861 SSIM, improving over the public Restormer $\sigma\!=\!50$ pretrained baseline by up to 3.366 dB PSNR. Ablation studies show that the dominant gain originates from the expanded training corpus and the two-stage optimization schedule, and self-ensemble provides marginal but consistent improvement.