Occluded Facial Recognition for Surviellance Using Deep Learning

Nowadays, due to the advancement in technology, facial recognition is becoming one of the methods to identify a person. One of the challenges arises due to occlusion or partial covering of face, especially with a facial mask or a scarf. In this work, we use deep neural networks to solve the problem of recognizing such an occluded face. For this work, we have used three publicly available facial datasets, namely Labelled Face Wild dataset, COMASK20 and Specs on Faces (with images having low illumination), cumulatively consisting more than 5000 facial images. We evaluated four existing facial detection classifiers namely OpenCV, Single Shot Detection(SSD), Multi-task Cascaded Convolutional Neural Network(MTCNN) and RetinaFace. We found MTCNN to be most relevant for our work. We proposed a new Convolutional Neural Networks (CNN) as part of this work. We got accuracy of 99.38% for LFW, 99.62% for COMASK20 and 98.33% for SOF dataset.

I. Introduction

Biometric recognition software is becoming more important in security, administration and business systems, the biometric systems can be finger print, iris recognition, retinal scanning, voice recognition and facial recognition. [1] [2] Facial recognition has sparked a lot of attention since it concentrates on detection, identification and verification. [3] It is now used in public places embedded with surveillance cameras, airport security and law enforcement agencies. Face recognition has been an important topic in scientific community for the past five years; nonetheless, it is still an unresolved problem that failed to recognize faces under occlusion. [4] [5] Occlusion can be due to multiple reasons like facial accessories such as eyeglasses, sunglasses, scarves, mask, hat, hair, and factors like extreme illumination.