Deep human face analysis and modelling

Human face appearance and motion play a significant role in creating the complex social environments of human civilisation. Humans possess the capacity to perform facial analysis and come to conclusion such as the identity of individuals, understanding emotional state and diagnosing diseases. The capacity though is not universal for the entire population, where there are medical conditions such prosopagnosia and autism which can directly affect facial analysis capabilities of individuals, while other facial analysis tasks require specific traits and training to perform well. This has lead to the research of facial analysis systems within the computer vision and machine learning fields over the previous decades, where the aim is to automate many facial analysis tasks to a level similar or surpassing humans. While breakthroughs have been made in certain tasks with the emergence of deep learning methods in the recent years, new state-of-the-art results have been achieved in many computer vision and machine learning tasks. Within this thesis an investigation into the use of deep learning based methods for facial analysis systems takes place, following a review of the literature specific facial analysis tasks, methods and challenges are found which form the basis for the research findings presented. The research presented within this thesis focuses on the tasks of face detection and facial symmetry analysis specifically for the medical condition facial palsy. Firstly an initial approach to face detection and symmetry analysis is proposed using a unified multi-task Faster R-CNN framework, this method presents good accuracy on the test data sets for both tasks but also demonstrates limitations from which the remaining chapters take their inspiration. Next the Integrated Deep Model is proposed for the tasks of face detection and landmark localisation, with specific focus on false positive face detection reduction which is crucial for accurate facial feature extraction in the medical applications studied within this thesis. Evaluation of the method on the Face Detection Dataset and Benchmark and Annotated Faces in-the-Wild benchmark data sets shows a significant increase of over 50% in precision against other state-of-the-art face detection methods, while retaining a high level of recall. The task of facial symmetry and facial palsy grading are the focus of the finals chapters where both geometry-based symmetry features and 3D CNNs are applied. It is found through evaluation that both methods have validity in the grading of facial palsy. The 3D CNNs are the most accurate with an F1 score of 0.88. 3D CNNs are also capable of recognising mouth motion for both those with and without facial palsy with an F1 score of 0.82