Research map / Label-efficient and robust learning / Domain adaptation and generalization
Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment
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 DGQA, which selects synthetic source domains (KADID-10k distortion types) whose distortion distribution is similar to an unlabeled authentic target via a multi-source domain classifier, then trains a BIQA ResNet-50 regressor on only the selected subset.
Problem
Synthetic-to-authentic generalization in blind image quality assessment is poor, and adding more synthetic distortion types can hurt transfer because outlier source domains cause negative transfer under unsupervised multi-source domain adaptation.
Approach
Filter multi-source synthetic distortion domains at the data level by scoring how often a distortion-type classifier assigns target images to each source domain, keeping only domains above a threshold (1/k), and train the IQA regressor on the retained subset, which is orthogonal to model-based BIQA methods.
Key results
- On KADID-10k→LIVEC, DGQA reaches SRCC 0.6958 and PLCC 0.6902, versus 0.5646/0.5810 for the full-data baseline, a 13.12% SRCC gain while using under 1/3 of the training data.
- Average SRCC over LIVEC, KonIQ-10k and BID is 0.7155 for DGQA versus 0.6404 for the best prior UDA method StyleAM.
- On the synthetic-to-algorithmic PIPAL setting, DGQA improves average SRCC from 0.4278 to 0.4718, with gains on five of six subtypes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select similar KADID-10k distortion domains instead of training on all domains (ablation) | DGQA (ResNet-50) · LIVEC | SRCC: 0.5646 -> 0.6958 (+13.12%) | Helped |
| Select similar KADID-10k distortion domains instead of training on all domains (ablation) | DGQA (ResNet-50) · LIVEC | PLCC: 0.5810 -> 0.6902 (+10.92%) | Helped |
| Select similar KADID-10k distortion domains instead of training on all domains (ablation) | DGQA (ResNet-50) · KonIQ-10k | SRCC: 0.6250 -> 0.6810 (+5.60%) | Helped |
| Select similar KADID-10k distortion domains instead of training on all domains (ablation) | DGQA (ResNet-50) · KonIQ-10k | PLCC: 0.6127 -> 0.6866 (+7.39%) | Helped |
| Add distortion-guided domain selection (ablation) | HyperIQA · LIVEC | SRCC: 0.5633 -> 0.6944 (+13.11%) | Helped |
| Add distortion-guided domain selection (ablation) | HyperIQA · KonIQ-10k | PLCC: 0.5869 -> 0.6628 (+7.59%) | Helped |
| Add distortion-guided domain selection (ablation) | StyleAM · LIVEC | SRCC: 0.5174 -> 0.6687 (+15.13%) | Helped |
| Add distortion-guided domain selection (ablation) | StyleAM · KonIQ-10k | SRCC: 0.7046 -> 0.7088 (+0.43%) | Helped |
| Use DGQA rather than StyleAM | KADID-10k-trained BIQA · KonIQ-10k | SRCC: 0.7002 -> 0.6810 | Hurt |
| Use DGQA rather than StyleAM | KADID-10k-trained BIQA · BID | SRCC: 0.6365 -> 0.7696 | Helped |
| Select similar KADID-10k distortion domains instead of using all domains | DGQA (ResNet-50) · PIPAL denoising | SRCC: 0.5705 -> 0.5588 (−1.17%) | Hurt |
| Select similar KADID-10k distortion domains instead of using all domains | DGQA (ResNet-50) · PIPAL PSNR-originated SR | SRCC: 0.4758 -> 0.5810 (+10.52%) | Helped |
Surprising findings
- Training on the larger synthetic dataset KADID-10k with more distortion types generalizes worse to LIVEC than the smaller LIVE dataset.
- A subset of KADID-10k source domains selected by greedy search generalizes to KonIQ-10k with higher SRCC than training on LIVEC itself, despite similar data magnitude.
Limitations
- Assumes source and target share the same content distribution, which is a theoretical assumption rather than verified.
- Denoising subtype on PIPAL is slightly worse than the baseline (SRCC -1.17%).
- Remaining gap to training on authentic datasets (e.g., KonIQ-10k to LIVEC) is still large.
Experimental setup
Datasets: KADID-10k KADIS-700k LIVEC BID KonIQ-10k PIPAL
Benchmarks: LIVEC KonIQ-10k BID PIPAL
Metrics: SRCC PLCC
Baselines: BRISQUE NIQE PIQE RankIQA MetaIQA DB-CNN HyperIQA DANN UCDA RankDA StyleAM
Builds on: DANN HyperIQA StyleAM DB-CNN
Closest papers by meaning
- Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions (NeurIPS 2025 · similarity 0.89)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.88)
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.87)
- Semi-Supervised Blind Quality Assessment with Confidence-quantifiable Pseudo-label Learning for Authentic Images (ICML 2025 · similarity 0.87)
- Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental Learning (AAAI 2024 · similarity 0.85)
- Enhancing Quality of Compressed Images by Mitigating Enhancement Bias Towards Compression Domain (CVPR 2024 · similarity 0.84)
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.