Research map / Image generation / Autoregressive and latent generation
Transformer-Based No-Reference Image Quality Assessment via Supervised Contrastive Learning
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 SaTQA, an NR-IQA model that pre-trains a ResNet-50 with supervised contrastive learning on the synthetic KADIS dataset to learn degradation features, then fuses them with Transformer-CNN multi-stream features via a patch attention block.
Problem
No-reference image quality assessment generalizes poorly to unknown authentic distortions, since existing CNN and Transformer models are trained without distinguishing distortion types and levels.
Approach
Use supervised contrastive learning on a large synthetic dataset (KADIS), with labels given by distortion type and level and no subjective scores, to learn degradation features. Combine these with perceptual features from a ViT-based multi-stream backbone (CNN plus Transformer branches) through cross-attention in a patch attention block.
Key results
- SaTQA reaches average PLCC 0.910 and SROCC 0.889 across seven datasets, versus 0.891 and 0.874 for the second-best MANIQA.
- On LIVE Challenge, SaTQA gets PLCC 0.903 and SROCC 0.877 (the paper reports these for its full model), and on LIVE-FB it reaches PLCC 0.676 against 0.625 for TReS.
- Ablation on LIVE Challenge: adding SCL raises PLCC from 0.827 to 0.871, and adding MSB and PAB brings it to 0.903.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SaTQA instead of the strongest listed baseline | NR-IQA · LIVE-FB | PLCC: 0.625 -> 0.676 | Helped |
| use SaTQA instead of the strongest listed baseline | NR-IQA · KONIQ | SROCC: 0.915 -> 0.930 | Helped |
| add supervised contrastive learning (SCL) (ablation) | SaTQA · LIVE Challenge | PLCC: 0.827 -> 0.871 | Helped |
| add supervised contrastive learning (SCL) (ablation) | SaTQA · CSIQ | PLCC: 0.922 -> 0.963 | Helped |
| add the Multi-Stream Block (MSB) (ablation) | SaTQA with SCL · LIVE Challenge | PLCC: 0.871 -> 0.880 | Helped |
| add the Multi-Stream Block (MSB) (ablation) | SaTQA with SCL · CSIQ | PLCC: 0.963 -> 0.970 | Helped |
| add Patch Attention Block (PAB) (ablation) | SaTQA with SCL and MSB · LIVE Challenge | PLCC: 0.880 -> 0.903 | Helped |
| add Patch Attention Block (PAB) (ablation) | SaTQA with SCL and MSB · CSIQ | PLCC: 0.970 -> 0.972 | Helped |
| use MSB rather than MR50 in the top backbone branch (ablation) | SaTQA · CSIQ | PLCC: 0.966 -> 0.972 | Helped |
| use MR50 rather than MSB in the bottom pre-training branch (ablation) | SaTQA · CSIQ | PLCC: 0.970 -> 0.972 | Helped |
| combine MHSA, DeformConv, and DwConv branches (MA+DF+DW) (ablation) | MSB · CSIQ | PLCC: 0.957 -> 0.972 | Helped |
| combine MHSA, DeformConv, and DwConv branches (MA+DF+DW) (ablation) | MSB · LIVE Challenge | PLCC: 0.878 -> 0.903 | Helped |
Surprising findings
- The paper reports that MSB alone does not improve over SCL on LIVE Challenge by more than 0.01, and that MSB-based pre-training is slower and more complex, with MR50 pre-training performing better than MSB in the heterogeneous-structure study.
Limitations
- Gains on synthetic datasets are small because results are near saturation.
- Performance on LIVE-FB remains low (PLCC 0.676), and the paper does not solve large-scale authentic distortion.
Experimental setup
Datasets: KADIS LIVE CSIQ TID2013 KADID-10K LIVE Challenge KONIQ LIVE-FB
Benchmarks: LIVE CSIQ TID2013 KADID-10K LIVE Challenge KONIQ LIVE-FB
Metrics: PLCC SROCC
Baselines: DIIVINE BRISQUE ILNIQE BIECON MEON WaDIQaM DBCNN TIQA MetaIQA P2P-BM HyperIQA TReS
Builds on: supervised contrastive learning ViT CONTRIQUE MANIQA CBAM
Closest papers by meaning
- Boosting Image Quality Assessment through Efficient Transformer Adaptation with Local Feature Enhancement (CVPR 2024 · similarity 0.89)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.88)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.87)
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (ICML 2024 · similarity 0.87)
- Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity (ICML 2024 · similarity 0.86)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · 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.