Cross polarimetric thermal-to-visible heterogeneous face recognition by coupled dictionary learning

Abstract

Cross-model heterogeneous face recognition (HFR) has been one of the most challenging areas of research in biometrics and computer vision. The main goal of HFR is to accurately recognize/identify the visible face image with probe images captured in alternative sensing modalities, such as thermal spectrum. The polarization state information of thermal faces contains the missing textural and geometrics details in the conventional thermal face imagery, which facilitate the development of cross-spectrum face recognition. In this paper, we propose a coupled dictionary learning architecture to find a common embedding space between the visible and sensing domains, where we model the learning problem as a bilevel optimization. The learned coupled dictionaries are used to transform visible and polarimetric thermal face images into the common embedding feature space via patch-wise sparse recovery. Experiments conducted with the polarimetric thermal facial datasets which contains face images that has been taken at three different ranges and with different face expressions. The results show that our proposed coupled learning method could fuse polarimetric and thermal features in a way to outperform the conventional methods and enhance the performance of a thermal-to-visible face recognition system.