Infrared Face Recognition Based On Deep Learning The University Of Chinese Academy Of Sciences

Infrared-based heterogeneous face recognition is a promising solution for identifying people in dark environments.Recent studies mainly focus on the near-infrared(NIR)to visible(VIS)face matching strategies,which maybe a sub-optimal solution for this task due to the reflective nature of NIR.The thermal-infrared(IR)based heterogeneous recognition provides the query images are collected from the radiation of facial skin.However,due to the expensive data collection process and large domain gap,the IR-VIS face recognition only receives limited attention in recent years.This paper focuses on infrared face recognition using deep learning technology,which can be summarized as follows:On one hand,we explore the network structure of face recognition(IR50,etc)and the optimization function of ArcFace Loss on several existing near-infrared datasets.With the IR50 network and Arc Face Loss,the recognition accuracy of our model is higher than 99%on Casia Nir-Vis 2.0 dataset.On the other hand,due to the lack of a complete thermal-infrared dataset,we collect and construct a thermal-infrared face database with over 1000 people,which is the largest thermal-infrared face dataset as we known.The samples of this database are collected under different distances and different temperatures.On this dataset,we perform a set of extensive experiments for thermal-infrared-to-visual heterogeneous face recognition task.Specifically,we discuss the challenge of this cross-modal database,and propose a Dual Alignment Neural Network(DANN)to reduce the domain gaps between two different modals.Experimental results show that our DANN achieves significant improvement on our ThermalFace database,which demonstrate the effectiveness of our proposed modules.We hope our baseline model can provide a promising direction for future work.