Closed-set thermal face recognition via the synthesis approach

Abstract

Thermal face recognition has various applications in security, video-surveillance, military and government industries, etc. In contrast to the visible-light domain case, there is a lack of sufficiently large databases of thermal-face images. Therefore, it is difficult to construct a good face recognition system directly through learning procedures. We address this problem by using the synthesis approach, in which artificial visible-domain images of faces are generated from thermal face images and then identified in the visible-light domain. The generation of the artificial face images is done using a conditional Generative Adversarial Network, and the face identification is done by another, separate, deep neural network. Generating visible face images from thermal ones using DNNs requires large datasets of pairs of thermal and visible face images of the same persons, taken from the same position, time and imaging conditions, with two different cameras. The available datasets of such pairs are of relatively very low amount of subjects. Therefore, as in other similar studies, we approach this problem as a closed-set face recognition task, in which all testing identities are predefined in the training set. We train the face identification network itself using the synthesized images together with the original ground-truth visible-domain images to create a recognition system that is suitable to the characteristics of the artificially generated images. Thus, the recognition network will be optimized specifically for the thermal face recognition.