Generalization Assessment of Palmprint Verification Models Trained on Synthetic Data Across Diverse Datasets

Palmprint verification systems have shown great potential as a reliable biometric modality due to their distinctiveness and stability. However, developing robust palmprint verification models requires large-scale and diverse training data, which is often difficult to acquire. This paper investigates the effectiveness of leveraging synthetic palmprint data for enhancing the performance and generalization capabilities of palmprint verification systems. We utilize the PCE-Palm, a novel two-stage synthetic palmprint generation method that transforms parametric Bézier curves into realistic palmprint images via an intermediate Palm Crease Energy (PCE) domain, to generate 1.6 million synthetic palmprint samples representing 50,000 subjects. Our approach involves pretraining deep learning models on synthetic data followed by fine-tuning on real palmprint datasets. We propose a Region of Interest (ROI) extraction method based on circular fitting for accurate palmprint region localization and employ EfficientNetV2 with Circle Loss for feature extraction and classification. Extensive experiments across ten diverse public palmprint datasets demonstrate that models pretrained on synthetic data significantly outperform models initialized with ImageNet weights. Specifically, our synthetic data-driven approach achieves superior results with Equal Error Rates (EER) as low as 0.0026 on COEP, 0.0045 on IITD, 0.0046 on TCD, and 0.0622 on PolyU_V2 datasets. These results highlight the potential of synthetic data for improving the generalization and robustness of palmprint verification systems across various acquisition conditions and population demographics.