On Designing a Full Palm Imaging and Recognition System

In palmprint recognition, biometric data collectors capture two separate palm impressions (upper and lower), using double the time and disk space for what could ideally be a singular whole palm image. The solution is to use a device that can electronically capture the entire palm at once, adding benefits including reduced chances of non-compliance. However, current devices typically lack speed, are costly, or lack sufficient resolution. This work discusses a novel palm scanning device (PalmLive) that captures the entire palm electronically in real time. Unlike document scanners, which are slow to produce images of similar quality in real time, PalmLive delivers them instantly. Additionally, the device’s cost is low compared to similarly capable commercial systems. Then, a new Palm+ dataset is collected, using a traditional palmprint, a document, and the PalmLive scanner. Next, a deep learning-based (DLB) distal detection model is developed that accurately detects distals in palm images. The work evaluates the recognition performance of the ground truth versus the detected palmprints. Experimental results showed verification accuracy improvements of 0.9%, 13.8%, and 16.7% with the model, for the traditional, document, and prototype scanners, respectively. The work concludes by assessing 20 different image quality algorithms and comparing their ability to separate good vs poor-quality distals for a palm quality assessment tool. The results show that the maximum correlation in Rank-1 accuracy of 0.904 was achieved by a non-biometrics DLB algorithm, while a recent biometric quality metric was superior for verification with a coefficient of 0.572.