Cross-device recognition of dorsal hand vein images by two-stage coarse-to-fine matching

This paper addresses the problem of poor biometric recognition performance that is caused by greatly increased variations among the hand vein patterns of the same individuals as a result of using different imaging devices under different acquisition conditions and presents a novel solution based on two-stage coarse-to-fine matching. In particular, a global pattern descriptor is proposed as a geometrical reference for optimum image segmentation of vein patterns without significant over-segmentation and under-segmentation. In order to accommodate large cross-device variations, a control parameter is introduced to allow adjustment of segmented vein patterns, thereby enabling not only intra-class pattern similarities but also inter-class pattern dissimilarities to be increased. Furthermore, overlapping of principal vein patterns is proposed as a criterion for global coarse matching to reduce the number of candidates for identification, and distinctive efficient robust features are employed to provide a biological vision-based descriptor of salient local pattern characteristics for fine matching. Using a large dataset of 2000 cross-device hand vein images captured from two different near-infrared imaging devices and 100 hands, the efficacy of the proposed approach for a cross-device biometric system is demonstrated, with a recognition performance shown to be compatible to that of a single-device biometric system.