Near Infrared Face Recognition using End to End Light Convolution Neural Network

Face Recognition in the visible spectrum is a widely used biometric tool. However, it is often influenced by environmental lighting that reduces its performance drastically. So as to solve the effects of illumination, face images are captured in the near-infrared (NIR) spectrum. While several deep learning models are available to solve the problem of face recognition in the visible spectrum, only a few models exist in the NIR spectrum. This paper proposes an end to end light CNN architecture that performs face recognition in the NIR spectrum. The proposed architecture has given higher accuracies for CASIA NIR VIS 2.0 (98.16%) and Oulu CASIA (99.26%) which are considered to be the largest and the most challenging publicly available datasets. Further, it has produced high accuracies for other publicly available datasets namely PolyU (97.90%), CBSR (98.22%) and HITSZ (99.72%).

I. Introduction

Face is considered as one of the most important biometric traits as it is contactless which has gathered high demand ever since COVID 19 pandemic. Face recognition (FR) is defined as the method by which the face images from a digital or a video source is used to identify/verify people. Over the last few decades, there has been a significant increase in demand for facial recognition systems as they are non-intrusive, contactless and are a cheaper technology for authentication. They have also shown remarkable advantages in forensics, surveillance and immigration. However, face images captured in the visible spectrum are highly sensitive to illumination variations. This problem can be solved by capturing the images in near-infrared spectrum (NIR). It provides an effective, low cost and straightforward solution to increase the efficiency of face recognition in poor lighting or in complete darkness. The face images taken in both near-infrared and visible spectrum under various lighting conditions such as strong, weak and dark illuminations are shown in Fig. 1. It can be observed that the face images do not vary under near-infrared with strong, weak and dark lighting conditions. Hence, working on NIR face images is widely adopted in mobile devices, video surveillance (night vision CCTV camera) and user authentication applications to solve the illumination effect. Fig. 1.

Visible and the corresponding NIR face images of the same subject in various (stong, weak and dark) illuminations from Oulu CASIA NIR VIS face dataset

TABLE I Details of Nir Face Datasets

Dataset	No. of Subjects	No. of Images
CASIA (NIR) 2.0	725	12,977
Oulu CASIA (NIR) facial expression	80	32,004
PolyU NIR	335	34,000
CBSR NIR	197	3,940
HITSZ NIR	50	500