Source-linked AI summary
ONOT/TONO-Print: A Paired Benchmark for Identity-Document Portrait Quality under Document Production Shift
Alexander Vinogradov
TL;DR
Face-quality models are commonly evaluated on digital images, although identity-document production can alter portraits’ visual properties. ONOT/TONO-Print provides paired synthetic and physical production variants for comparison; production effects vary across models, and the benchmark covers only a limited set of production conditions.
Problem
Identity-document production can change portrait appearance without changing underlying face information, raising the question of how face-quality models behave under this production shift.
Source: §1 · p. 1Method
Results
Production effects are model-dependent: some face-quality methods remain relatively stable, while others show substantial score changes under the same transformations.
Source: §4 · p. 2Takeaways & Limitations
The paired benchmark supports direct comparison of quality scores and component-level performance before and after identity-document production.
Source: §4 · p. 2Coverage is limited to five simulated production techniques and a small physical inkjet/toner subset, and does not reproduce the full complexity of document manufacturing, substrates, laminates, or security elements.
Source: §4 · p. 2
Abstract
from arXiv · showhide
Identity-document portraits are modified during document production by printing, engraving, and integration into the document template. These transformations may change image texture, contrast, colour, and other visual properties without necessarily changing the underlying biometric or presentation information. In this work, we introduce ONOT/TONO-Print, a small paired benchmark for evaluating face-image quality models under identity-document production shift. The benchmark is based on ONOT and TONO portraits and contains synthetic variants representing five common production techniques: toner, inkjet, dye sublimation, laser engraving, and thermal transfer. A smaller physical subset contains portraits printed and scanned using inkjet and toner devices. The benchmark is used to compare the behaviour of biometric and ICAO-related face-quality components before and after production. Results show that production can introduce substantial domain shift for some quality estimators, while other models remain comparatively stable.
1. Introduction
Identity-document production can alter portrait appearance without changing the underlying face information, potentially affecting face-quality measures. ONOT/TONO-Print adds controlled production transformations and a physical subset to study model robustness under this shift.
1. Introduction
- Identity-document portraits may be printed, engraved, or transferred to physical documents and combined with document-specific security and design elements.Source: §1 · p. 1
- Production can change texture, contrast, colour, and sharpness, so quality models may respond to production effects rather than changes in face information.Source: §1 · p. 1
- ONOT/TONO-Print extends controlled portrait data with production transformations and a small physical print-and-scan subset to study robustness under production-related domain shift.Source: §1 · p. 1
2. Dataset description
ONOT/TONO-Print pairs ONOT and corresponding TONO portraits with synthetic production variants and a smaller physical print-and-scan subset. The transformations approximate production effects while preserving source identity and presentation condition.
2. Dataset description
- The benchmark uses portraits from 256 ONOT identities and corresponding TONO portraits.Source: §2 · p. 1
- Five synthetic production transformations approximate characteristic effects while preserving each source portrait’s identity and presentation condition.Source: §2 · p. 1, Figure 1 · p. 1
- Synthetic examples show the digital source alongside toner, inkjet, dye sublimation, laser engraving, and thermal-transfer variants.Source: Figure 1 · p. 1, §2 · p. 1
- A smaller physical subset uses real inkjet and toner printing followed by scanning as a sanity check, not as coverage of all manufacturing technologies.Source: §2 · p. 1, Figure 2 · p. 1
- Each transformed sample is linked to its corresponding digital source using the original ONOT identity naming.Source: §2 · p. 2
3. Experiments
The experiments compare five face-quality models using digital portraits as references for production sensitivity. The reported effect of production varies substantially by model.
3. Experiments
- The evaluation includes OFIQ UQS, CR-FIQA, MagFace, eDifFIQA, and FaceQAN.Source: §3 · p. 2
- Production effects are strongly model-dependent in the reported relative score changes.Source: §3 · p. 2, Table 2 · p. 2
- For each method, production sensitivity is measured as the change from the digital portrait reference to the transformed condition.Source: §3 · p. 2
4. Conclusions and Future Work
ONOT/TONO-Print enables paired comparisons of face-quality scores and component-level performance before and after production. Results vary across models, while the benchmark remains limited in production coverage.
4. Conclusions and Future Work
- Production-related degradation is not reflected consistently across face-quality models: some remain relatively stable, while others show substantial score changes under the same transformations.Source: §4 · p. 2
- The benchmark separates original portrait content from production transformations for direct before-and-after comparison of quality scores and component-level performance.Source: §4 · p. 2
- Coverage is limited to five simulated techniques and a small physical inkjet/toner subset, without the full complexity of manufacturing, substrates, laminates, or security elements.Source: §4 · p. 2
- Future work may add physical production techniques and document types.Source: §4 · p. 2