Unconstrained biometric recognition based on thermal hand images

This paper proposes a biometric recognition method based on thermal images of inner part of the hand, and a database of 21,000 thermal images of both hands acquired by a specialized thermal camera from 70 subjects. The data for each subject was acquired in three different sessions, with two first sessions organized on the same day, and the third session organized approximately two weeks apart. This allowed to analyze the stability of hand temperature in both short-term and long-term horizons. No hand stabilization or positioning devices were used during acquisition, making this setup closer to real-world, unconstrained applications. This required making our method translation-, rotation- and scale-invariant. Two approaches for feature selection and classification are proposed and compared: feature engineering deploying texture descriptors such as Binarized Statistical Image Features (BSIF) and Gabor wavelets, and feature learning based on convolutional neural networks (CNN) trained in different environmental conditions. For within-session scenario we achieved 0.36% and 0.00% of equal error rate (EER) in the first and the second approach, respectively. Between-session EER stands at 27.98% for the first approach and 17.17% for the second one. These results allow for estimation of a short-term stability of hand thermal information. This paper presents the first known to us database of hand thermal images and the first biometric system based solely on hand thermal maps acquired by thermal sensor in unconstrained scenario.