Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-Supervised Blind Quality Assessment with Confidence-quantifiable Pseudo-label Learning for Authentic Images
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 CPL-IQA, a single-branch semi-supervised blind IQA method that converts MOS to vector labels via entropy minimization and iteratively learns confidence-weighted pseudo-labels via label propagation on a kNN graph.
Problem
Blind image quality assessment for authentic distortions has few labeled images, and existing semi-supervised BIQA methods need extra branches or auxiliary data, while label propagation was designed for classification and scalar MOS labels.
Approach
Convert scalar MOS labels into 100-dim vector labels via entropy minimization so label propagation can be applied to regression, then learn pseudo-label vectors for unlabeled images using manifold-based label optimization and weight them by an entropy-based confidence, all in a single ResNet101 branch trained end-to-end.
Key results
- On KonIQ-10K (1:3:1), CPL-IQA reaches PLCC 0.873 and SRCC 0.845, slightly above SSLIQA (0.867/0.841) and Semi-IQA (0.864/0.839).
- In cross-data tests trained on KonIQ-10K, CPL-IQA reaches PLCC 0.777 on LIVE-C (vs. 0.706 for SSLIQA) and 0.772 on NNID.
- Ablation on SPAQ (1:8:1) shows Label-only PLCC 0.812 rising to 0.896 with full CPL-IQA including confidence weights.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Convert MOS labels to vector labels with entropy minimization (ablation) | CPL-IQA · KonIQ-10K (1:3:1) | PLCC: 0.713 -> 0.850 | Helped |
| Convert MOS labels to vector labels with entropy minimization (ablation) | CPL-IQA · SPAQ (1:8:1) | PLCC: 0.812 -> 0.873 | Helped |
| Add Stage 2 label optimizing and pseudo-label training (ablation) | CPL-IQA Stage 1 · KonIQ-10K (1:3:1) | PLCC: 0.850 -> 0.862 | Helped |
| Add confidence weighting for pseudo-labels (ablation) | CPL-IQA · KonIQ-10K (1:3:1) | PLCC: 0.862 -> 0.873 | Helped |
| Add confidence weighting for pseudo-labels (ablation) | CPL-IQA · SPAQ (1:8:1) | PLCC: 0.881 -> 0.896 | Helped |
| Add confidence weighting for pseudo-labels (ablation) | CPL-IQA · SPAQ (2:7:1) | PLCC: 0.898 -> 0.903 | Helped |
| Increase score-set cardinality from m=10 to m=100 (ablation) | CPL-IQA · KonIQ-10K (1:3:1) | PLCC: 0.850 -> 0.873 | Helped |
| Increase score-set cardinality from m=10 to m=100 (ablation) | CPL-IQA · SPAQ (1:8:1) | PLCC: 0.884 -> 0.896 | Helped |
| Increase labeled-sample share from split 1:7:2 to 3:5:2 (ablation) | CPL-IQA · KonIQ-10K (m=20) | PLCC: 0.845 -> 0.875 | Helped |
| Increase labeled-sample share from split 1:8:1 to 3:6:1 (ablation) | CPL-IQA · SPAQ (m=20) | PLCC: 0.892 -> 0.905 | Helped |
| Use CPL-IQA instead of Causal-IQA | BIQA models trained on KonIQ-10K · LIVE-C (cross-data) | PLCC: 0.742 -> 0.777 | Helped |
| Use CPL-IQA instead of Re-IQA | BIQA models trained on KonIQ-10K · NNID (cross-data) | SRCC: 0.770 -> 0.773 | Helped |
Surprising findings
- Traditional label propagation designed for classification performs poorly on IQA regression, so cosine-similarity graphs were replaced with raw-feature kNN graphs.
Limitations
- Gains over the strongest semi-supervised baselines on KonIQ-10K are small (e.g., 0.873 vs 0.867 PLCC).
- Label conversion via entropy minimization assumes a one-to-one MOS-to-distribution mapping without prior information, which the authors admit may not hold.
Experimental setup
Datasets: KonIQ-10K LIVE-C NNID SPAQ BID KADID-10K ImageNet
Benchmarks: KonIQ-10K LIVE-C NNID SPAQ
Metrics: PLCC SRCC KRCC RMSE
Baselines: NIQE BRISQUE GWH-GLBP SSEQ CNNIQA WaDIQaM PAQ-2-PIQ NSSADNN GraphIQA MB-CNN MUSIQ Causal-IQA
Builds on: label propagation NIMA EMD loss kNN graph Label-Only (Fang et al., 2020)
Closest papers by meaning
- Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental Learning (AAAI 2024 · similarity 0.90)
- Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment (CVPR 2024 · similarity 0.87)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.86)
- Transformer-Based No-Reference Image Quality Assessment via Supervised Contrastive Learning (AAAI 2024 · similarity 0.85)
- Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions (NeurIPS 2025 · similarity 0.85)
- Contrastive Credibility Propagation for Reliable Semi-supervised Learning (AAAI 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.