Robust Cross-Spectral Periocular Biometric Recognition using LoRA-Adapted DINOv2 and Quality-Aware Fusion

Abstract

Robust periocular recognition in contactless settings remains challenging due to cross-sensor and cross-spectral appearance variations, as well as unconstrained factors such as pose, blur, and occlusion. This paper proposes a parameter-efficient periocular recognition framework based on a DINOv2-pretrained foundation Vision Transformer (ViT) adapted using Low-Rank Adaptation (LoRA). The LoRA modules are injected into the query, key, and value projections of the transformer backbone, while an ArcFace head is employed to learn identity-discriminative embeddings. During enrollment, multiple gallery samples of each subject are encoded and combined using a Quality-Aware Gated Fusion (QAGF) strategy to construct a reliability-aware identity template, where informative samples receive higher contribution and degraded samples are automatically down-weighted. The proposed framework is evaluated on two benchmark periocular datasets, UBIPr and UTIRIS, under visible-spectrum, near-infrared-spectrum, and cross-spectrum protocols. Ablation studies involving frozen backbone transfer, full fine tuning, and LoRA-based parameter-efficient fine-tuning further validate the effectiveness of the proposed adaptation strategy in terms of both recognition performance and trainable parameter reduction. Experimental results demonstrate consistent improvements over CNN baselines and representative prior methods, achieving an Equal Error Rate (EER) as low as 1.2% and a True Accept Rate (TAR) of 99.1% at False Accept Rate (FAR) = 10−3, while reducing trainable parameters by more than 90%. These results indicate that LoRA-adapted foundation vision transformers with quality-aware template fusion provide an effective and scalable solution for next-generation periocular biometric recognition.