Research map / Image generation / Autoregressive and latent generation
Integrating Global Context Contrast and Local Sensitivity 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 CSIQA, a BIQA method that combines quality-label-based contrastive learning (with easy/hard sampling and a distillation teacher) for global dataset context with a quality-aware mask attention token for local distortion sensitivity.
Problem
No-reference image quality assessment often treats samples independently and relies on semantic pre-trained features, missing cross-image quality contrast and local distortion sensitivity.
Approach
Defines contrastive positives and negatives by ground-truth quality similarity, refines them with easy/hard sampling using teacher pseudo-label errors, and adds a quality-aware (QUA) token whose attention is restricted to randomly unmasked patches to focus on local distortions.
Key results
- CSIQA reports PLCC/SRCC of 0.941/0.926 on TID2013 and 0.920/0.898 on LIVEC, outperforming prior SOTA methods on six of eight datasets.
- Plugging quality context contrastive learning into TIQA and DEIQT improves them by roughly 1 to 3 points on CSIQ, LIVEC and KonIQ.
- Cross-dataset SRCC on LIVEC and KonIQ improves over DEIQT and HyperIQA (e.g., 0.787 on LIVEC when trained on KonIQ).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CSIQA instead of DEIQT | BIQA · LIVE | PLCC: 0.982 -> 0.985 | Helped |
| Use CSIQA instead of CONTRIQUE | BIQA · KADID | PLCC: 0.937 -> 0.919 | Hurt |
| Use CSIQA instead of Re-IQA | BIQA · LIVEFB | PLCC: 0.733 -> 0.688 | Hurt |
| Use CSIQA instead of Re-IQA | BIQA · SPAQ | PLCC: 0.925 -> 0.935 | Helped |
| Add quality context contrastive learning and distillation (ablation) | TIQA · CSIQ | PLCC: +1.8 | Helped |
| Add quality context contrastive learning and distillation (ablation) | TIQA · LIVEC | PLCC: +2.9 | Helped |
| Add quality context contrastive learning and distillation (ablation) | DEIQT · CSIQ | PLCC: +0.7 | Helped |
| Add quality context contrastive learning and distillation (ablation) | DEIQT · LIVEC | SRCC: +1.8 | Helped |
| Add mask attention (ablation) | CSIQA · TID2013 | PLCC | Helped |
| Add easy-sample contrastive learning (ablation) | CSIQA · TID2013 | PLCC: approximately 1.7% | Helped |
| Add hard sampling to easy-sample contrastive learning (ablation) | CSIQA · TID2013 | PLCC: nearly 1% | Helped |
Surprising findings
- Authentic images with different content but similar quality are treated as negatives by prior distortion-based contrastive methods, which the authors argue is harmful.
Limitations
- Random masking introduces noise into attention that may need further regularization.
- Performance gains vary across datasets, attributed to batch size and positive/negative sample scarcity.
- Gains on SPAQ and LIVEFB are not shown as best; CSIQA does not lead on all datasets.
Experimental setup
Datasets: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Benchmarks: LIVE CSIQ TID2013 KADID LIVEC KonIQ LIVEFB SPAQ
Metrics: SRCC PLCC
Baselines: BRISQUE ILNIQE BIECON MEON WaDIQaM DBCNN TIQA MetaIQA P2P-BM HyperIQA TReS MUSIQ
Builds on: InfoNCE DEIQT MUSIQ Masked autoencoders
Closest papers by meaning
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.91)
- Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental Learning (AAAI 2024 · similarity 0.88)
- Feature Denoising Diffusion Model for Blind Image Quality Assessment (AAAI 2025 · similarity 0.88)
- Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment (CVPR 2024 · similarity 0.87)
- KVQ: Boosting Video Quality Assessment via Saliency-guided Local Perception (CVPR 2025 · similarity 0.86)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · 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.