Research map / Continual learning / Online and replay-based
Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental 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 SS-IQA, a semi-supervised blind image quality assessment framework that combines KRR-based knowledge distillation for pseudo-labels with incremental learning and SDK-Sample experience replay to avoid forgetting.
Problem
Blind image quality assessment needs large labeled datasets, which are often scarce; unsupervised methods perform poorly, and supervised fine-tuning causes forgetting of the source domain.
Approach
Treats unlabeled target data as incremental blocks: a teacher trained on labeled data produces pseudo-labels via kernel ridge regression over features, the student is trained on strongly augmented blocks, and a smallest-difference top-K replay mechanism selects exemplars to prevent catastrophic forgetting.
Key results
- On LIVE->CSIQ, SS-IQA reaches PLCC 0.969 / SRCC 0.960 on CSIQ while keeping LIVE at 0.980/0.978 (only -0.002 SRCC drop relative to labeled baseline).
- On CSIQ->LIVE, the unlabeled-target SS-IQA achieves PLCC 0.964 / SRCC 0.949 on LIVE, outperforming unsupervised methods such as ILNIQE (0.814/0.879) and NIQE.
- Ablation on LIVE->CSIQ shows baseline 0.957/0.940 rising to 0.964/0.949 on CSIQ with full SS-IQA (incremental learning and KRR both contribute).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add incremental learning (ablation) | SS-IQA · LIVE → CSIQ | PLCC: 0.957 -> 0.963 | Helped |
| Add KRR-based knowledge distillation (ablation) | SS-IQA · LIVE → CSIQ | PLCC: 0.957 -> 0.961 | Helped |
| Combine incremental learning and KRR-based distillation (ablation) | SS-IQA · LIVE → CSIQ | PLCC: 0.957 -> 0.964 | Helped |
| Add incremental learning (ablation) | SS-IQA · KADID → TID2013 | PLCC: 0.918 -> 0.927 | Helped |
| Add KRR-based knowledge distillation (ablation) | SS-IQA · KADID → TID2013 | PLCC: 0.918 -> 0.921 | Helped |
| Combine incremental learning and KRR-based distillation (ablation) | SS-IQA · KADID → TID2013 | PLCC: 0.918 -> 0.929 | Helped |
| Vary the number of unlabeled-data blocks from Q=1 to Q=6 (ablation) | SS-IQA · CSIQ → LIVE | PLCC: 0.938, 0.943, 0.948, 0.937, 0.952 | Mixed |
| Vary the number of unlabeled-data blocks from Q=1 to Q=6 (ablation) | SS-IQA · KADID → TID2013 | PLCC: 0.925, 0.928, 0.929, 0.929, 0.931 | Mixed |
| Vary representative-sample frequency K1 and K2 (ablation) | SS-IQA · LIVE → CSIQ | PLCC: 0.977, 0.978, 0.979, 0.979, 0.979, 0.979, 0.979, 0.977, 0.978, 0.980 | No clear effect |
Surprising findings
- On the authentic-to-authentic tasks (LIVEFB->SPAQ, and LIVEC->KonIQ-related settings), performance on the labeled source domain decreases and the method is weaker than some supervised baselines.
- Varying the number of DB blocks Q and replay sample sizes K1, K2 changes performance by less than about 1%, suggesting low sensitivity.
Limitations
- Performance on the LIVEFB->SPAQ and similar authentic tasks is comparatively poor; the method does not always improve across incremental phases when unlabeled data lies outside the original data space.
- Relies on a teacher trained on a labeled source dataset, so requires related labeled data to transfer from.
Experimental setup
Datasets: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ ImageNet-1K
Benchmarks: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Metrics: PLCC SRCC
Baselines: ILNIQE MEON WaDIQaM DBCNN TIQA MetaIQA P2P-BM HyperIQA TReS MUSIQ DACNN SSLIQA
Builds on: DeiT III HyperIQA Grad-CAM
Closest papers by meaning
- Semi-Supervised Blind Quality Assessment with Confidence-quantifiable Pseudo-label Learning for Authentic Images (ICML 2025 · similarity 0.90)
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.88)
- Feature Denoising Diffusion Model for Blind Image Quality Assessment (AAAI 2025 · similarity 0.87)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.87)
- Knowledge Guided Semi-supervised Learning for Quality Assessment of User Generated Videos (AAAI 2024 · similarity 0.85)
- Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment (CVPR 2024 · similarity 0.85)
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.