Learning Face Similarities for Face Verification using Hybrid Convolutional Neural Networks

• Published in: Indonesian Journal of Electrical Engineering and Computer Science
• Main Author: Kamaru Zaman F.H.; Johari J.; Yassin A.I.M.
• Format: Article
• Language: English
• Published: Institute of Advanced Engineering and Science 2019
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85073550318&doi=10.11591%2fijeecs.v16.i3.pp1333-1342&partnerID=40&md5=3e355d881f1d9b96797605b7a038d954

Face verification focuses on the task of determining whether two face images belong to the same identity or not. For unrestricted faces in the wild, this is a very challenging task. Besides significant degradation due to images that have large variations in pose, illumination, expression, aging, and occlusions, it also suffers from large-scale ever-expanding data needed to perform one-to-many recognition task. In this paper, we propose a face verification method by learning face similarities using a Convolutional Neural Networks (ConvNet). Instead of extracting features from each face image separately, our ConvNet model jointly extracts relational visual features from two face images in comparison. We train four hybrid ConvNet models to learn how to distinguish similarities between the face pair of four different face portions and join them at top-layer classifier level. We use binary-class classifier at top-layer level to identify the similarity of face pairs which includes a conventional Multi-Layer Perceptron (MLP), Support Vector Machines (SVM), Native Bayes, and another ConvNet. There are 3 face pairing configurations discussed in this paper. Results from experiments using Labeled face in the Wild (LFW) and CelebA datasets indicate that our hybrid ConvNet increases the face verification accuracy by as much as 27% when compared to individual ConvNet approach. We also found that Lateral face pair configuration yields the best LFW test accuracy on a very strict test protocol without any face alignment using MLP as top-layer classifier at 87.89%, which on-par with the state-of-the-arts. We showed that our approach is more flexible in terms of inferencing the learned models on out-of-sample data by testing LFW and CelebA on either model. Copyright © 2019 Institute of Advanced Engineering and Science. All rights reserved.