Improving Low-Resolution Face Recognition under Limited Data: How Synthetic Data Generation Can Close the Domain Gap
Date:
Fifteen-minute oral presentation in the Focus Session on Generative AI for Fair and Secure Biometrics under Limited Data, plus a poster.
Surveillance face recognition often has to match faces of only 16–32 pixels against high-resolution galleries, and paired low/high-resolution training data is scarce. We studied synthetic low-resolution data generation for a compact face recognition model (EdgeFace-S) at three levels of effort: interpolation, Real-ESRGAN-style degradation, and a learned identity-aware super-resolution front-end. Main findings:
- The degradation that is best on synthetic benchmarks is the worst on real low-resolution faces (TinyFace).
- More synthesis effort does not pay off monotonically: simple interpolation augmentation beats the more complex options on the compact model.
- Feeding the aligned low-resolution face directly to a strong backbone is a hard baseline that the super-resolution pipeline does not beat.
- Accuracy gains do not reduce demographic bias on RFW.
Authors: Luis S. Luevano, Ünsal Öztürk, Hatef Otroshi Shahreza, Anjith George, Sébastien Marcel.