Face Recognition in the Dark: A Unified Approach for NIR- VIS and VIS- NIR Face Matching

Visible face recognition security systems fail to recognize the faces in total darkness. Near Infrared (NIR) spectrum provides an affordable and effectual method to capture good quality images in poor lighting or complete darkness condition. In this paper, a unified model has been proposed for both NIRVIS and VIS- NIR heterogeneous face recognition to solve the illumination problem. The proposed model is based on transfer learning which is a modified version of the Resnet-34 model and it avoids training on large scale NIR face images. Two experiments have been conducted using the proposed model on publicly available datasets, namely Oulu- CASIA and HITSZ. NIR query images are matched with VIS gallery images whereas, in the second experiment, VIS query images are matched with NIR gallery images. The proposed model has given a rank-l accuracy of 98.54% on the Oulu- CASIA dataset which outperforms the existing state-of-the-art models.