Research map / Probabilistic and scientific modeling / Causal discovery
Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference
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 a training-free full-reference IQA method that uses causal interventions and a causal transport cost on pretrained VGG/ResNet/EfficientNet deep features, evaluated across six IQA benchmarks.
Problem
Full-reference image quality assessment methods rely on feature correlations and statistical dependencies, failing to capture causal mechanisms by which distortions affect human perceived quality, which limits generalization across datasets and distortion types.
Approach
Frames FR-IQA as abductive counterfactual inference: applies do-interventions with an exogenous variable on reference and distorted deep features, keeps features whose perceptual distance changes under intervention (via a confounder dictionary), and measures perceptual cost with a causal transport distance, without training.
Key results
- Our-VGG reaches 0.929 PLCC / 0.932 SRCC on LIVE and 0.949 / 0.952 on CSIQ, outperforming the listed baselines on those datasets.
- On PIPAL, Our-VGG scores 0.675 PLCC / 0.657 SRCC, below TOPIQ-FR (0.837 / 0.809, trained on PIPAL) and DeepWSD-free baselines such as PieAPP (0.702 / 0.701).
- Ablation on Our-VGG shows the causal-feature weights (phi_gamma) give 0.909 / 0.884 PLCC/SRCC on TID2013 versus 0.884 / 0.867 for original pretrained weights.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace original pretrained weights with causal-effect weights (ablation) | Our-VGG · LIVE | PLCC: 0.901 -> 0.929 | Helped |
| replace original pretrained weights with causal-effect weights (ablation) | Our-VGG · CSIQ | PLCC: 0.913 -> 0.949 | Helped |
| replace original pretrained weights with causal-effect weights (ablation) | Our-VGG · TID2013 | PLCC: 0.884 -> 0.909 | Helped |
| use causal-effect weights instead of residual weights excluding causal effects (ablation) | Our-VGG · LIVE | PLCC: 0.843 -> 0.929 | Helped |
| use causal-effect weights instead of residual weights excluding causal effects (ablation) | Our-VGG · CSIQ | PLCC: 0.803 -> 0.949 | Helped |
| use causal-effect weights instead of residual weights excluding causal effects (ablation) | Our-VGG · TID2013 | PLCC: 0.786 -> 0.909 | Helped |
| replace original pretrained weights with causal-effect weights (ablation) | Our-VGG · LIVE | SRCC: 0.915 -> 0.932 | Helped |
| replace original pretrained weights with causal-effect weights (ablation) | Our-VGG · CSIQ | SRCC: 0.916 -> 0.952 | Helped |
| replace original pretrained weights with causal-effect weights (ablation) | Our-VGG · TID2013 | SRCC: 0.867 -> 0.884 | Helped |
| use causal-effect weights instead of residual weights excluding causal effects (ablation) | Our-VGG · LIVE | SRCC: 0.866 -> 0.932 | Helped |
| use causal-effect weights instead of residual weights excluding causal effects (ablation) | Our-VGG · CSIQ | SRCC: 0.831 -> 0.952 | Helped |
| use causal-effect weights instead of residual weights excluding causal effects (ablation) | Our-VGG · TID2013 | SRCC: 0.789 -> 0.884 | Helped |
Surprising findings
- On PIPAL, the proposed training-free method is outperformed by PieAPP and TOPIQ-FR, which were trained on that dataset, so it does not dominate across all benchmarks.
Limitations
- Method is training-free and its causal formulation is loosely specified; key quantities (confounder dictionary, intervention procedure) are described informally.
- Underperforms trained methods such as TOPIQ-FR and PieAPP on PIPAL.
- Causal claims are supported mainly by ablations and qualitative visualizations rather than formal causal validation.
Experimental setup
Datasets: LIVE CSIQ TID2008 TID2013 KADID-10k PIPAL
Benchmarks: LIVE CSIQ TID2008 TID2013 KADID-10k PIPAL
Metrics: PLCC SRCC
Baselines: PSNR SSIM MS-SSIM VIF MAD FSIM VSI GMSD NLPD PieAPP LPIPS DISTS
Builds on: LPIPS DISTS DeepWSD
Closest papers by meaning
- Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal Inference (ICML 2024 · similarity 0.90)
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.86)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.86)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.85)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.85)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · 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.