Loss Functions for CNN-based Biometric Vein Recognition

The recent progress in deep learning has led to a rapid change in the way biometric data can be handled, offering new opportunities for further research on physical, behavioral, and cognitive biometric recognition. In particular, conventional modalities for preprocessing, extracting features, and comparing templates derived from biometric traits have been swiftly altered, replacing the search for hand-crafted features with the ever-increasing use of generalized deep learning models and transfer learning, able to guarantee notably-high recognition performance. This study investigates the capabilities of deep learning approaches in performing vein pattern verification. Specifically, recent advances in the design of convolutional neural networks, introduced to increase the inter-class variability and decrease the intra-class variability of the generated representations, are here taken into account to speculate on the effects on recognition performance of the selection for the most suitable loss function. Experimental tests conducted on finger vein, palm vein, and hand dorsum vein patterns testify the effectiveness of the proposed frameworks, able to exceed current state-of-the-art performance on five different publicly available vein datasets.