Research map / Image generation / Image restoration
DSL-FIQA: Assessing Facial Image Quality via Dual-Set Degradation Learning and Landmark-Guided Transformer
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 DSL-FIQA, a transformer-based generic face image quality assessment model with self-supervised dual-set degradation learning and landmark guidance, plus a new 40K-image CGFIQA-40k dataset.
Problem
Generic face image quality assessment (GFIQA) of perceptual quality, where prior patch-based degradation learning assumes uniform degradation within an image and existing datasets are small or demographically imbalanced.
Approach
Learns degradation representations globally by contrasting a synthetic-degradation set (same content, different degradations) against a real-world GFIQA set via soft proximity mapping and bidirectional contrastive loss, injects them into a transformer decoder via cross-attention, and uses positional-encoded facial landmark identifiers to guide a regional confidence map.
Key results
- DSL-FIQA reaches PLCC 0.9745 / SRCC 0.9740 on GFIQA-20k, outperforming StyleGAN-IQA (0.9673/0.9684).
- On CGFIQA-40k DSL-FIQA achieves PLCC 0.9873 / SRCC 0.9880, above the best baseline ReIQA (0.9800/0.9802).
- DSL raises real-world degradation retrieval mAP from 48.2 (patch-based) to 69.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace patch-based degradation learning with DSL (ablation) | GFIQA network · GFIQA-20k | PLCC: 0.9695 -> 0.9721 | Helped |
| Use cross-attention instead of concatenating degradation representations (ablation) | DSL-FIQA · GFIQA-20k | PLCC: 0.9703 -> 0.9721 | Helped |
| Use bidirectional DSL loss instead of one-sided DSL-S loss (ablation) | DSL-FIQA · GFIQA-20k | PLCC: 0.9713 -> 0.9721 | Helped |
| Use bidirectional DSL loss instead of one-sided DSL-R loss (ablation) | DSL-FIQA · GFIQA-20k | PLCC: 0.9709 -> 0.9721 | Helped |
| Add landmark guidance (ablation) | DSL-FIQA · GFIQA-20k | PLCC: 0.9721 -> 0.9731 | Helped |
| Add positional encoding to landmark identifiers (ablation) | DSL-FIQA with landmark guidance · GFIQA-20k | PLCC: 0.9731 -> 0.9735 | Helped |
| Use Charbonnier loss (ablation) | DSL-FIQA · GFIQA-20k | PLCC: 0.9735 -> 0.9745 | Helped |
Limitations
- Off-the-shelf landmark detectors fail on heavily degraded images, requiring fine-tuning.
- Evaluated only on face IQA datasets; generalization beyond faces not shown.
- Image is cropped into patches and MOS averaged, which may limit global context.
Experimental setup
Datasets: GFIQA-20k PIQ23 CGFIQA-40k FFHQ
Benchmarks: GFIQA-20k PIQ23 CGFIQA-40k
Metrics: PLCC SRCC mAP
Baselines: StyleGAN-IQA ReIQA MANIQA TRIQ IFQA CONTRIQUE MUSIQ HyperIQA
Builds on: StyleGAN-IQA ReIQA MANIQA NeRF positional encoding
Closest papers by meaning
- CLIB-FIQA: Face Image Quality Assessment with Confidence Calibration (CVPR 2024 · similarity 0.88)
- Overcoming False Illusions in Real-World Face Restoration with Multi-Modal Guided Diffusion Model (ICLR 2025 · similarity 0.85)
- Toward Generalized Image Quality Assessment: Relaxing the Perfect Reference Quality Assumption (CVPR 2025 · similarity 0.84)
- Transformer-Based No-Reference Image Quality Assessment via Supervised Contrastive Learning (AAAI 2024 · similarity 0.84)
- Self-Supervised Selective-Guided Diffusion Model for Old-Photo Face Restoration (NeurIPS 2025 · similarity 0.84)
- Learning Degradation-unaware Representation with Prior-based Latent Transformations for Blind Face Restoration (CVPR 2024 · similarity 0.84)
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.