Research map / Label-efficient and robust learning / Domain adaptation and generalization

Uncertainty-Aware Source-Free Adaptive Image Super-Resolution with Wavelet Augmentation Transformer

CVPR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation · 10 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 SODA-SR, the first source-free domain adaptation framework for real-world image super-resolution, combining a teacher-student setup with a wavelet augmentation transformer, uncertainty-aware pseudo-label rectification, and frequency regularization losses.

Problem

Adapting a source-trained image super-resolution model to a target domain (synthetic-to-real or cross-camera real-to-real) using only unlabeled target LR images, without access to source data, since source data may be unavailable due to privacy or transmission restrictions.

Approach

A teacher-student framework with EMA teacher, 8-fold geometric self-ensembling for refined pseudo-labels, Gumbel-Softmax-based MC sampling for uncertainty-weighted pseudo-label loss, and a Wavelet Augmentation Transformer (batch attention over wavelet low-frequency sub-bands plus deformable cross-level attention) that is used only in training as implicit augmentation.

Key results

• On DRealSR real-to-real adaptation (Panasonic->Sony), SODA-SR reports 32.24 PSNR, the best among compared methods, without source data, versus DADA's 32.13.
• On synthetic-to-real settings, SODA-SR beats UDA baselines across all six camera pairs (e.g., 31.61 vs 31.50 PSNR for Panasonic->Sony).
• Ablation on Olympus->Panasonic: removing WAT drops PSNR from 31.41 to 30.64, and removing EMA drops it to 31.14.

Measured effects reported in the paper

Surprising findings

• The source-free SODA-SR outperforms source-accessing UDA methods such as DADA on PSNR and SSIM in most camera pairs.

Limitations

• Evaluated only on DRealSR with a single 4x scale and three camera pairs.
• Pseudo-label quality is still limited when the source-target domain gap is large, as the method relies on the source model's initial predictions.

Experimental setup

Datasets: DRealSR

Benchmarks: DRealSR

Metrics: PSNR SSIM LPIPS

Baselines: CinCGAN DASR DRN-Adapt DADA ZSSR MZSR Source Only Target Only

Builds on: CDC DADA Gumbel-Softmax mean teacher deformable attention

Closest papers by meaning

• Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.86)
• IODA: Instance-Guided One-shot Domain Adaptation for Super-Resolution (NeurIPS 2024 · similarity 0.86)
• AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · similarity 0.85)
• Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts (CVPR 2024 · similarity 0.85)
• AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.85)
• Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach (CVPR 2025 · similarity 0.85)