Dermatologically inspired deep face analytics

Abstract

This PhD thesis explores the use of convolutional neural networks (CNNs) for the extraction of age related health information from face images. CNNs are well suited to face image analysis thanks to their ability to learn robust features from large diverse datasets. Our research focuses on illuminating the intersection between deep learning for face image analysis and healthy ageing, leveraging the key methods supporting state-of-the-art age estimation. We also investigate the influence of image preprocessing on face analysis, emphasising the significance of face alignment algorithms and other more novel techniques. Finally, we validate the summation of our work using health outcomes, the gold standard endpoint for measures of ageing. The first part of our study revolves around image preprocessing techniques, investigating the influence of face alignment algorithms in downstream face image analysis tasks and proposing our own system for eye based alignment. Additionally, we investigate the efficacy of semi-supervised methods to facilitate learning from a novel medical dataset containing 3D rendered faces with heavy confounding artefacts. The second objective is to analyse the properties of different CNN based methods for perceived age estimation. We assess the strengths and weaknesses of transfer learning schemes, layer con- figurations and unsupervised learning approaches. Through extensive experimentation, we elu- cidate the optimal design choices and achieve a new state-of-the-art in CNN-based perceived age estimation. Lastly we assess the usefulness of our perceived age models on their association with relevant health outcomes and genetic markers, providing objective measures of its efficacy, free from the bias of human perception. We demonstrate the potential of automated perceived age estimations linked to health outcomes, paving the way for accelerated biological research by removing the bottleneck of human assessors.