Research map / Image generation / Image restoration
Perceptual Fairness in Image Restoration
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 Perceptual Fairness, defined via a Group Perceptual Index (distributional divergence between each group's ground truth and reconstructions), and shows it detects bias that Representation Demographic Parity misses in face super-resolution.
Problem
Existing fairness definitions for image restoration (RDP, PR, CPR) treat a reconstruction as correct only if it lies in the group's ground-truth support, ignoring distributional quality and graded errors, so they can miss group disparities.
Approach
Define Group Perceptual Index as a statistical distance between each group's ground-truth and reconstruction distributions, and call an algorithm perceptually fair when GPIs are equal across groups; prove relations to group precision/recall and the perception-distortion/PI tradeoff.
Key results
- Group PI via KID reveals a clear disadvantage for old&Asian versus old&non-Asian, while RDP-style group precision is similar (roughly zero GP difference) between these groups.
- Adversarial age-attack on non-Asian inputs of RestoreFormer++ is undetected by group precision (RDP) but clearly identified by GPI-KID of each group.
- Theory: perfect GPI for all groups is infeasible under sufficiently severe degradation, and perfect PI with a majority group implies imperfect PF under TV distance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| evaluate groups with GPI-KID instead of group precision (GP) | face super-resolution algorithms · face super-resolution; old&Asian vs. old&non-Asian | GPI-KID | Helped |
| use DINOv2 or CLIP features instead of FairFace features to compute GPI-KID (ablation) | face super-resolution algorithms · synthetic face groups | GPI-KID group disparity | Mixed |
| use Inception-v3 features instead of FairFace features to compute GPI-KID (ablation) | face super-resolution algorithms · old&Asian vs. old&non-Asian | GPI-KID group comparison | Mixed |
| adversarially attack non-Asian inputs to make restored faces appear 70+ | RestoreFormer++ · face super-resolution; young&non-Asian group | GPI-KID | Hurt |
| adversarially attack non-Asian inputs to make restored faces appear 70+ | RestoreFormer++ · face super-resolution; young&non-Asian group | GP | No clear effect |
| use GPI-KID rather than GPSNR or GLPIPS to assess group fairness | face super-resolution algorithms · young&non-Asian vs. young&Asian | group bias detection | Helped |
Surprising findings
- Under perfect perceptual index, improving one group's GPI can improve other groups' GPI even when groups do not overlap.
- GPSNR and GLPIPS show roughly similar values across groups where GP and GPI-KID reveal clear disparities, so distortion metrics appear unreliable bias indicators.
Limitations
- Theorems are preliminary and the optimal estimators achieving good PF are not characterized.
- Requires a sensitive-attribute classifier for feature extraction; general-purpose feature extractors (DINOv2, CLIP) reveal less pronounced biases and Inception features give inconsistent results.
- Evaluated only on human face images with synthetic, generator-produced fairness groups that may carry biases and stereotypes.
Experimental setup
Datasets: CelebA-HQ FairFace
Benchmarks: CelebA-HQ synthetic fairness groups (old/young x Asian/non-Asian)
Metrics: Group Perceptual Index (KID) Group Perceptual Index (FID) Group Precision (classification hit rate) Group PSNR Group LPIPS Group precision/recall (NN)
Baselines: DDNM+ DDRM DPS PiGDM GFPGAN VQFR GPEN DiffBIR CodeFormer RestoreFormer++ RestoreFormer
Builds on: Representation Demographic Parity Perception-distortion tradeoff Kernel Inception Distance FairFace
Closest papers by meaning
- Do Counterfactually Fair Image Classifiers Satisfy Group Fairness? -- A Theoretical and Empirical Study (NeurIPS 2024 · similarity 0.85)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.85)
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2024 · similarity 0.83)
- FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in Medical Image Analysis (ICLR 2024 · similarity 0.83)
- The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective (NeurIPS 2025 · similarity 0.83)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.83)
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.