Research map / Probabilistic and scientific modeling / Causal discovery
Causal-IQA: Towards the Generalization of Image Quality Assessment Based on Causal 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 Causal-IQA, a causality-based meta-learning training paradigm for blind image quality assessment that uses counterfactual division over distortion types and GMM-derived content clusters with backdoor adjustment to improve cross-dataset generalization.
Problem
Deep learning blind IQA models trained on small synthetic IQA datasets overfit and generalize poorly to unseen distortions and authentic images, because distortion type and image content confound the mapping from image to MOS.
Approach
Builds a structural causal graph for IQA, forms sample subsets with fixed distortion type and content cluster (Counterfactual Division), and approximates backdoor adjustment P(m|do(I)) via a Reptile-style meta-learning update over these subsets (CIS serial, CIP parallel variants).
Key results
- On authentic datasets after fine-tuning, CIS reaches SROCC 0.895 on CID2013, 0.828 on LIVE-C and 0.881 on KonIQ-10K, versus MetaIQA's 0.766, 0.802 and 0.850.
- In zero-shot testing on authentic images (Table 4), CIS and CIP improve PLCC/SROCC by about 7-8 points over MetaIQA.
- Leave-one-distortion-out on KADID-10K gives average SROCC of 0.8078 (CIS) and 0.7974 (CIP) versus 0.7672 for MetaIQA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add image-content confounder handling to CIS (ablation) | CIS · CID2013, LIVE-C, and KonIQ-10K zero-shot | SROCC: 0.642 -> 0.651; 0.652 -> 0.684; 0.628 -> 0.658 | Helped |
| add image-content confounder handling to CIP (ablation) | CIP · CID2013, LIVE-C, and KonIQ-10K zero-shot | SROCC: 0.641 -> 0.683; 0.615 -> 0.624; 0.613 -> 0.704 | Helped |
| use serial rather than parallel task-loss updating (ablation) | Causal-IQA · CID2013, LIVE-C, and KonIQ-10K fine-tuning | PLCC: 0.873 -> 0.887; 0.844 -> 0.847; 0.896 -> 0.918 | Mixed |
| use serial rather than parallel task-loss updating (ablation) | Causal-IQA · CID2013, LIVE-C, and KonIQ-10K zero-shot | SROCC: 0.683 -> 0.651; 0.624 -> 0.684; 0.704 -> 0.658 | Mixed |
| increase backbone depth from ResNet18 to ResNet50 (ablation) | CIS · CID2013 | PLCC: 0.887 -> 0.921 | Helped |
| increase backbone depth from ResNet18 to ResNet50 (ablation) | CIS · LIVE-C | PLCC: 0.847 -> 0.885 | Helped |
| increase backbone depth from ResNet18 to ResNet50 (ablation) | CIS · KonIQ-10K | PLCC: 0.918 -> 0.936 | Helped |
| increase GMM cluster number beyond 2 (ablation) | Causal-IQA · LIVE-C fine-tuning | PLCC and SROCC | No clear effect |
| increase fine-tuning training ratio (ablation) | Causal-IQA · LIVE-C fine-tuning | PLCC and SROCC | Helped |
| use CIS rather than MetaIQA | ResNet18 IQA model · CID2013 | PLCC: 0.784 -> 0.887 | Helped |
| use CIS rather than MetaIQA | ResNet18 IQA model · LIVE-C | SROCC: 0.802 -> 0.828 | Helped |
| use CIS rather than MetaIQA | ResNet18 IQA model · KonIQ-10K | PLCC: 0.887 -> 0.918 | Helped |
Surprising findings
- CIS (serial updating) outperforms CIP (parallel updating) in most settings, and the content-cluster component is not strictly necessary for some gains (CIS\C already beats baseline by a wide margin).
Limitations
- Only distortion type is modeled as a confounder; distortion degree is left to the appendix.
- Content clusters are obtained by unsupervised GMM on ResNet18 features with c=2 chosen, since content labels are unavailable.
- Causal-IQA is evaluated only on synthetic training sets with ResNet backbones; gains on larger-scale or transformer-based models are not shown.
Experimental setup
Datasets: TID2013 KADID-10K KonIQ-10K LIVE-C CID2013
Benchmarks: TID2013 KADID-10K KonIQ-10K LIVE-C CID2013
Metrics: PLCC SROCC
Baselines: BLIINDS-II BRISQUE ILNIQE CORNIA HOSA BIECON MEON WaDIQaM-NR DIQA NSSADNN MetaIQA DistNet-Q3
Builds on: MetaIQA Reptile MAML ResNet
Closest papers by meaning
- Image Quality Assessment: Investigating Causal Perceptual Effects with Abductive Counterfactual Inference (CVPR 2025 · similarity 0.90)
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.88)
- Towards Syn-to-Real IQA: A Novel Perspective on Reshaping Synthetic Data Distributions (NeurIPS 2025 · similarity 0.88)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.88)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.87)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (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.