An In-Depth Analysis of Face Recognition Models: A Comparative Study of Deep Learning Architectures on the Labeled Faces in the Wild Dataset

Abstract

 Artificial Intelligence is one of the important tools widely used in our society mainly Deep Learning (DL) which has quite a lot of applications due to its ability to learn robust representations from images and videos for recognition tasks. Convolutional Neural Networks (CNN) is a subset of DL heavily used by researchers following the breakthrough of AlexNet by winning the most difficult image classification standard ImageNet Large Scale Visual Recognition Challenge (ISLVRC) 2012 by decreasing the error by 10% from the winning algorithm of the 2011 on the benchmark. This paper compares the performance of ten face recognition models built using DL architectures on Labelled Faces in Wild (LFW) dataset: DeepFace, FaceNet and VGGFace were trained using triplet loss which have lower convergence but achieved higher accuracy. DeepID, DeepID2, DeepID2+ and DeepID3 were trained using sofmax loss, but learn features that are not discriminative enough because the linear transformation matrix's size grows the number of identities increases linearly, SphereFace reached good performance but the training process is unstable, ArcFace attained the state of art performance by introducing Additive Angular Margin Loss to mitigate the two main problems associated with the previous approaches.