Detecting Thermal Face Signature Abnormalities

Abstract

In this paper, we propose a novel method of applying deep learning techniques to face biometrics in infrared spectrum. It addresses detection of abnormal thermal patterns, thus identifying, in particular, indicators of insobriety. This finds its application for security and healthcare emergency detection in city shelters. We applied the deep learning approach on 16,000 usable images of 40 subjects from a publicly available Drunk-Sober database. Two Convolutional Neural Network architectures were investigated for the task of processing of two regions of interest - the forehead and the eyes. The accuracy of the neural network classifiers to predict subject's insobriety using the forehead and eye regions-of-interest reached 95.5% and 96.67%, respectively, compared to to best known results on the same data using a non-deep neural networks. To boost the accuracy of classification, both the feature-level and the score-level fusion were applied as well, thus improving the accuracy to 96.92%.