Beyond the Visible Spectrum: Is Person Identity Well Preserved in Thermal Cameras?

Abstract

AbstractFace recognition is a well investigated problem in the computer vision community. Beyond the visible spectrum, it is identified as an active-oriented and attractive research area. In particular, thermal cameras have recently emerged as increasingly important sensors for visual surveillance applications. In addition to the fact that these cameras operate well in challenging environments such as adverse weather and lighting conditions, they are commonly known as keystone biometric solution that preserves person identity. In this paper, we intend to prove that faces could be highly recognized from thermal cameras using a powerful generative adversarial model. This model is employed to deal with the domain shift between thermal and visible sensors. Extensive experiments of different generative models and face recognition systems demonstrate the effectiveness of the proposed pipeline to reveal the person identity even though it is acquired by different sensing modalities, with significant facial variations.KeywordsGenerative Adversarial NetworkThermalVisibleFace recognitionCross-spectral