Face recognition by thermal video using 3D information and vesselness features

This paper aims to introduce a new approach for face recognition in thermal imagery by exploiting the depth information hidden in infrared video. The process starts by reconstructing a 3D model of the head from this video which contains frames in different poses to provide a depth image. On the other hand, the original thermal images which describe the intensity information can be enhanced and be transformed to a vascular network which is proven to be an efficient anatomical feature for face recognition. These two informations are regrouped to form an image where each pixel has two values, one for depth data and another for vessel intensity. This image goes through a hierarchical feature selection system using linear discriminant analysis and an Adaboost method learning to produce a feature template and to construct a final strong classifier. Numerical results on real date highlight the relevance of our proposed approach.