Research map / Image generation / Image restoration
Distilling Spatially-Heterogeneous Distortion Perception 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 SHDIQA, a ViT-based blind IQA model that trains on block-wise degraded images with distortion tokens and distills contrastive and affinity knowledge from a teacher to capture spatially heterogeneous distortions.
Problem
Blind image quality assessment of authentically distorted images whose distortions vary spatially, which global whole-image degradation priors fail to model.
Approach
Applies distinct randomly chosen distortions to different image blocks to create local distortion priors, uses per-block distortion (DIS) tokens with attention masks alongside a CLS token, and distills contrastive (CKD) and affinity (AKD) knowledge from a quality-pretrained teacher to model inter-block distortion relations.
Key results
- SHDIQA (24.2M params) achieves average PLCC/SRCC of 0.984/0.982 on LIVE, 0.972/0.962 on CSIQ, 0.940/0.937 on KADID, and 0.735/0.639 on LIVEFB-style authentic benchmarks, outperforming listed SOTA methods on all seven datasets.
- Combining CKD and AKD improves the baseline by 6.2 points, and block-wise degradation outperforms homogeneous whole-image heterogeneous distortion (LIVEC SRCC 0.909 vs 0.893).
- With only 20% of training images on LIVEC, SHDIQA reaches PLCC 0.866 / SRCC 0.840, above LoDa (0.854/0.815) and DEIQT (0.822/0.792).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use spatially heterogeneous local degradation and distortion-aware aggregation and distillation | SHDIQA · LIVEFB | SRCC: 0.578 -> 0.639 | Helped |
| add contrastive knowledge distillation (ablation) | SHDIQA · LIVEC | SRCC: 0.900 -> 0.909 | Helped |
| add affinity knowledge distillation (ablation) | SHDIQA · LIVEC | SRCC: 0.867 -> 0.909 | Helped |
| add distortion classification supervision (ablation) | SHDIQA · KADID | SRCC: 0.884 -> 0.904 | Helped |
| add teacher-output regularization (ablation) | SHDIQA · KADID | SRCC: 0.884 -> 0.904 | Helped |
| replace L1 with BCE contrastive loss (ablation) | SHDIQA CKD · LIVEC | SRCC: 0.895 -> 0.902 | Helped |
| apply affinity distillation to both original and degraded images (ablation) | SHDIQA AKD · LIVEC | SRCC: 0.901 -> 0.909 | Helped |
| use 4 degraded blocks instead of 3 (ablation) | SHDIQA · LIVEC | PLCC: 0.920 -> 0.922 | Helped |
| use 4 degraded blocks instead of 5 (ablation) | SHDIQA · LIVEC | PLCC: 0.918 -> 0.922 | Helped |
| add block-wise filtering after block-wise aggregation (ablation) | SHDIQA · LIVEC | SRCC: 0.902 -> 0.909 | Helped |
| apply distortions locally rather than heterogeneously over the whole image (ablation) | SHDIQA · LIVEC | SRCC: 0.893 -> 0.909 | Helped |
| replace the proposed contrastive loss with L1 loss (ablation) | SHDIQA CKD · LIVEC | SRCC: 0.902 -> 0.895 | Hurt |
Surprising findings
- Applying CKD alone gives only limited gains, which the authors attribute to aggressively separating regions that may share similar quality.
Limitations
- Gains on some benchmarks (e.g., LIVEFB) are reported with limited detail, and cross-dataset performance is not uniformly best (wins 4 of 6 cross-dataset settings).
- Block-wise degradation can cause content loss in degraded blocks, and the number of blocks trades off degradation scope against content loss.
Experimental setup
Datasets: LIVE CSIQ KADID LIVEC KonIQ LIVEFB SPAQ
Benchmarks: LIVE CSIQ KADID LIVEC KonIQ LIVEFB SPAQ
Metrics: SRCC PLCC
Baselines: BRISQUE DBCNN TIQA MetaIQA HyperIQA TReS MUSIQ Re-IQA DEIQT QPT LIQE QFM-IQM
Builds on: DEIQT Re-IQA QPT MoCo-v2 DeiT III
Closest papers by meaning
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.91)
- Bridging the Synthetic-to-Authentic Gap: Distortion-Guided Unsupervised Domain Adaptation for Blind Image Quality Assessment (CVPR 2024 · similarity 0.88)
- Transformer-Based No-Reference Image Quality Assessment via Supervised Contrastive Learning (AAAI 2024 · similarity 0.88)
- Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions (NeurIPS 2025 · similarity 0.88)
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (ICML 2024 · similarity 0.87)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · 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.