Face Recognition using Deep Learning in unconstrained environment

Abstract

Biometric identifiers are generally classified into physiological and behavioural characteristics. The shape of the body is considered in physiological characteristics which include palm print, finger print, iris recognition, face recognition etc. Face recognition has caught the attention of most of the researchers in the recent years as it has advantages over the other identifiers, as it does not require co-operation from the human and availability of many public datasets. The current face recognition methods have achieved results which are on par or better than human performance in constrained environment. However, their performances in unconstrained environment where the face images captured vary in illumination, occlusion, pose, resolution etc. is not satisfactory. In this work we propose to develop a face recognition model which is robust to those variations. As Convolutional neural network provide the state of art results in image classification, we develop a deep neural network using CNN. In order to meet the goal we created a dataset by downloading the images from the web and placing a standalone camera in the campus to collect the images from the videos captured. OpenCV is used to detect the face images. Our model developed with eight convolutional layers and trained on our dataset outperforms fine-tuned VGGFace model and Alexnet trained from the scratch. Our model with 5,671,824 parameters as compared to VGGFace fine-tuned, Alexnet 40,487,824 and 28,159,832 parameters respectively achieved test accuracy of 99.75% as compared to fine-tuned VGGFace, Alexnet which achieved 99.64% and 98.12% respectively. The proposed model occupies 68MB which is very less compared to fine-tuned VGGFace with 293MB and Alexnet with 338MB. newline