Research map / Image generation / Autoregressive and latent generation
Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity
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 QFM-IQM, an NR-IQA method that matches semantically dissimilar but similar-quality images as adversarial feature noise and enforces quality-prediction consistency, plus distillation-based pseudo-labeling of extra unlabeled images.
Problem
No-reference image quality assessment built on semantic backbones suffers from feature aliasing: images with similar semantic features but very different quality scores get confused, because upstream features contain quality-irrelevant semantic noise.
Approach
Treat images with similar quality scores but dissimilar semantic features as adversarial semantic noise (selected via a feature-similarity times score-distance metric from a temporary memory), mix them into the ViT features, and enforce consistent quality predictions with the ground-truth supervision, so the model becomes insensitive to semantic noise; a teacher pseudo-labels extra unlabeled image
Key results
- QFM-IQM reaches SRCC 0.981 on LIVE, 0.954 on CSIQ, 0.916 on TID2013 and 0.891 on LIVEC, outperforming the listed baselines on those datasets.
- Cross-dataset validation: trained on LIVEFB, tested on KonIQ, SRCC 0.768 vs 0.758 for HyperIQA and 0.733 for DEIQT.
- Ablation on TID2013: the full model (SNM + QCC on both labeled and unlabeled data) reaches SRCC 0.916 vs 0.891 for the baseline without these components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add QCC on labeled data (ablation) | QFM-IQM with SNM · LIVEC | PLCC: 0.883 -> 0.904 | Helped |
| Add QCC on labeled data (ablation) | QFM-IQM with SNM · TID2013 | PLCC: 0.899 -> 0.919 | Helped |
| Add QCC on unlabeled data (ablation) | QFM-IQM with SNM · TID2013 | PLCC: 0.899 -> 0.925 | Helped |
| Use QCC on both labeled and unlabeled data (ablation) | QFM-IQM with SNM · TID2013 | PLCC: 0.899 -> 0.932 | Helped |
| Match samples with similar quality and dissimilar features (ablation) | QFM-IQM · LIVEC | PLCC: 0.883 -> 0.898 | Helped |
| Match samples with similar quality and dissimilar features (ablation) | QFM-IQM · TID2013 | PLCC: 0.899 -> 0.916 | Helped |
| Use both quality-similar and feature-dissimilar matching (ablation) | QFM-IQM with QCC · LIVEC | PLCC: 0.883 -> 0.913 | Helped |
| Use LIVEFB as unlabeled data instead of no unlabeled data (ablation) | QFM-IQM DLE · TID2013 | PLCC: 0.919 -> 0.932 | Helped |
| Use KADID as unlabeled data instead of no unlabeled data (ablation) | QFM-IQM DLE · TID2013 | PLCC: 0.919 -> 0.921 | Helped |
| Use KonIQ as unlabeled data instead of no unlabeled data (ablation) | QFM-IQM DLE · TID2013 | PLCC: 0.919 -> 0.928 | Helped |
| Set memory search K to 6 rather than K=1 (ablation) | QFM-IQM · LIVEC | PLCC: 0.910 -> 0.913 | Helped |
| Increase memory search K from 6 to 8 (ablation) | QFM-IQM · LIVEC | PLCC: 0.913 -> 0.903 | Hurt |
Surprising findings
- QCC applied only to the labeled set A reduced stability on TID2013 (higher variance), which the authors attribute to synthetic datasets having less semantic richness.
- Re-IQA beats QFM-IQM on LIVEFB (SRCC 0.645 vs 0.567 for QFM-IQM in the first table, PLCC 0.733 vs 0.667), so the method is not best on all datasets.
Limitations
- Results are averaged over random 80/20 splits repeated ten times, which may not match standard benchmark protocols and make comparisons with baselines less direct.
- Performance is sensitive to hyperparameters lambda1, lambda2 and K, and K=1 vs K=6 optimum differs between synthetic and authentic datasets.
- Synthetic datasets offer limited semantic diversity, so the QCC constraint can reduce stability there.
Experimental setup
Datasets: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Benchmarks: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Metrics: SRCC PLCC
Baselines: DIIVINE BRISQUE ILNIQE BIECON MEON WaDIQaM DBCNN MetaIQA P2P-BM HyperIQA TReS MUSIQ
Builds on: DEIQT DeiT III TReS Re-IQA
Closest papers by meaning
- Exploring Semantic Feature Discrimination for Perceptual Image Super-Resolution and Opinion-Unaware No-Reference Image Quality Assessment (CVPR 2025 · similarity 0.88)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.88)
- Defense Against Adversarial Attacks on No-Reference Image Quality Models with Gradient Norm Regularization (CVPR 2024 · similarity 0.87)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.87)
- Guardians of Image Quality: Benchmarking Defenses Against Adversarial Attacks on Image Quality Metrics (ICML 2025 · similarity 0.87)
- Boosting Image Quality Assessment through Efficient Transformer Adaptation with Local Feature Enhancement (CVPR 2024 · similarity 0.87)
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.