IJCB 2026 study: simple synthetic data beats complex SR for low-res face recognition
Learned super-resolution failed to outperform simple interpolation in a new low-res face recognition study.
Face recognition systems in surveillance often encounter low-resolution faces—those below the standard 112×112 pixel input. While high-resolution (HR) training data is plentiful, labeled native low-resolution (LR) data and paired LR/HR data are scarce. A common workaround is synthesizing LR data from available HR faces, but it's unclear how much synthesis effort actually pays off in recognition accuracy. The team compared five strategies for a compact, edge-device-oriented system: interpolation-based degradation, knowledge distillation, a Prepended Domain Transformer (PDT), Real-ESRGAN-style degradation, and a learned super-resolution (SR) front-end with identity-aware loss.
Their findings reveal a critical synthetic-real gap: the degradation setting that works best on synthetic benchmarks (LFW, CFP-FP, AgeDB-30) isn't optimal on real LR data like TinyFace. More synthesis effort doesn't help monotonically—the learned SR front-end failed to surpass a direct feed of the aligned LR image into a strong backbone. Only simple interpolation augmentation of a compact backbone improved over its own baseline. The authors conclude that generative methods for LR face recognition must be validated on real LR data against a direct-feed baseline, and they released the pipeline for reproducibility.
- Compared 5 synthetic LR generation strategies, including Real-ESRGAN degradation and a learned SR front-end with identity-aware loss.
- Only simple interpolation augmentation improved accuracy over baseline on real-world TinyFace dataset.
- Learned super-resolution did not beat direct feeding of LR images into a strong backbone (synthetic-real gap).
Why It Matters
Shows that expensive generative models for low-res face recognition may be unnecessary—simple, efficient methods can work better on real surveillance data.