Design and Real-World Application of Novel Machine Learning Techniques for Improving Face Recognition Algorithms

Abstract

Recent progress in machine learning has made possible the development of real-world face recognition applications that can match face images as good as or better than humans. However, several challenges remain unsolved. In this PhD thesis, some of these challenges are studied and novel machine learning techniques to improve the performance of real-world face recognition applications are proposed.

Current face recognition algorithms based on deep learning techniques are able to achieve outstanding accuracy when dealing with face images taken in unconstrained environments. However, training these algorithms is often costly due to the very large datasets and the high computational resources needed. On the other hand, traditional methods for face recognition are better suited when these requirements cannot be satisfied. This PhD thesis presents new techniques for both traditional and deep learning methods. In particular, a novel traditional face recognition method that combines texture and shape features together with subspace representation techniques is first presented. The proposed method is lightweight and can be trained quickly with small datasets. This method is used for matching face images scanned from identity documents against face images stored in the biometric chip of such documents. Next, two new techniques to increase the performance of face recognition methods based on convolutional neural networks are presented. Specifically, a novel training strategy that increases face recognition accuracy when dealing with face images presenting occlusions, and a new loss function that improves the performance of the triplet loss function are proposed. Finally, the problem of collecting large face datasets is considered, and a novel method based on generative adversarial networks to synthesize both face images of existing subjects in a dataset and face images of new subjects is proposed. The accuracy of existing face recognition algorithms can be increased by training with datasets augmented with the synthetic face images generated by the proposed method. In addition to the main contributions, this thesis provides a comprehensive literature review of face recognition methods and their evolution over the years.

A significant amount of the work presented in this PhD thesis is the outcome of a 3-year-long research project partially funded by Innovate UK as part of a Knowledge Transfer Partnership between University of Hertfordshire and IDscan Biometrics Ltd (partnership number: 009547).