Enhancing Face Recognition Robustness Via Knowledge Distillation with Low-Resolution-Aware Teachers

Low-resolution (LR) face recognition remains challenging in unconstrained imagery, where faces often appear as small, degraded crops affected by blur, compression, and misalignment. Many practical deployments require lightweight models with low latency and memoryconstraints, which tend to degrade more strongly under LR inputs. This thesis investigates whether distilling from LR-aware teachers provides more useful supervision to a MobileFaceNet (MFN) student than distilling from standard high-performing teachers. Under a controlled framework with a fixed student and distillation objective, we compare standard CNNandtransformerteacherstoLR-specialized teachers that incorporate quality adaptive margins, alignment-robust training, or keypoint-relative spatial priors. We evaluate the resulting MFN students on TinyFace (LR-native identification) and on IJB B/IJB-C template-based verification at strict operating points, reported as TAR at fixed FAR (TAR@FAR). Overall, both the teacher’s robustness mechanism and the distillation objective influence LR performance, where margin-aligned distillation is more consistent than naive feature regression, and LR-aware teachers are most beneficial when training scale and augmentation match the teacher’s priors. Under WebFace4M, LR-aware teachers deliver the strongest LR gains. The MobileFaceNet baseline achieves 55.82%/61.93% TinyFace Rank-1/Rank-5, while ViT-B+KPRPE with teacher-aligned augmentation reaches 66.14%/70.94% (+10.32/+9.01 pp) and maintains 89.51% TAR at strict FAR=10-5 on IJB-C. Across matched backbones, LR-aware supervision improves over standard teacher distillation by up to +2.68/+2.09 pp (Rank-1/Rank-5), highlighting that teacher specialization matters for LR-FR distillation. We also observe a robustness-separability trade-off: stronger LR augmentation can increase TinyFace accuracy but may reduce performance at strict FAR, whereas KPRPE stabilizes this behavior.