Face Quality Assessment¶
Definition¶
Face Quality Assessment (FQA) evaluates whether a captured face image meets the minimum quality requirements for reliable biometric processing. Poor quality inputs lead to false rejections, failed matches, and unreliable liveness decisions.
Quality Factors¶
| Factor | Metric | Good | Marginal | Poor |
|---|---|---|---|---|
| Sharpness | Laplacian variance | > 200 | 50-200 | < 50 |
| Brightness | Mean luminance | 80-200 | 40-80 or 200-240 | < 40 or > 240 |
| Contrast | Standard deviation | > 40 | 20-40 | < 20 |
| Face size | Inter-ocular distance | > 90px | 60-90px | < 60px |
| Yaw angle | Degrees | < 15° | 15-30° | > 30° |
| Pitch angle | Degrees | < 15° | 15-25° | > 25° |
| Eye visibility | EAR score | Both open | One partially closed | Closed/occluded |
| Occlusion | Coverage % | < 5% occluded | 5-15% | > 15% |
| Expression | Neutral score | Neutral | Mild expression | Extreme expression |
ICAO Compliance (Passport/Travel Document Photos)¶
ICAO (International Civil Aviation Organization) defines strict quality standards:
| Requirement | ICAO Standard |
|---|---|
| Head position | Frontal, neutral expression |
| Eyes | Open, clearly visible, no red-eye |
| Mouth | Closed |
| Glasses | Preferably none; if worn, no reflections, eyes fully visible |
| Head covering | Only for religious reasons; face fully visible |
| Background | Plain, light, uniform |
| Lighting | Even, no shadows on face |
| Resolution | Minimum 300 DPI |
| Color | Full color (not B&W) |
Quality-Aware Processing¶
graph TD
A[Face Image] --> B[Quality Assessment]
B --> C{Quality Score}
C -->|High quality > 0.8| D[Standard processing<br/>Normal thresholds]
C -->|Medium quality 0.5-0.8| E[Adaptive processing<br/>Adjusted thresholds]
C -->|Low quality < 0.5| F[Reject + retry<br/>Guide user to improve]
E --> G[Lower match threshold<br/>Compensate with stronger liveness]
style D fill:#2E7D32,color:#fff
style F fill:#e53935,color:#fff
style E fill:#F57F17,color:#000Quality-Aware Models¶
| Model | Approach |
|---|---|
| AdaFace | Adapts loss function based on image quality during training |
| SER-FIQ | Self-supervised quality estimation from face embeddings |
| MagFace | Uses embedding magnitude as quality indicator |
| FaceQnet | Dedicated quality estimation network |
Key Takeaways¶
Summary
- Quality assessment is a critical gate — prevents poor images from entering the pipeline
- Key factors: sharpness, brightness, face size, pose angle, eye visibility, occlusion
- ICAO standards define the gold standard for face image quality
- Quality-aware models (AdaFace, MagFace) adapt their processing based on input quality
- Real-time quality feedback in the SDK dramatically improves first-attempt success rate