Face recognition operating in visible domains exists in many aspects of our lives, while the remaining parts of the spectrum including near and thermal infrared are not sufficiently explored. Thermal-visible face recognition is a promising biometric modality that combines affordable technology and high imaging qualities in the visible domain with low-light capabilities of thermal infrared. In this work, we present the results of our study in the field of thermal-visible face verification using four different algorithm architectures tested using several publicly available databases. The study covers Siamese, Triplet, and Verification Through Identification methods in various configurations. As a result, we propose a triple triplet face verification method that combines three CNNs being used in each of the triplet branches. The triple triplet method outperforms other reference methods and achieves TAR @FAR 1% values up to 90.61%.
Keywords: CNN; biometrics; cross-spectral face recognition; thermal to visible face recognition.