This present research is concerned with the design, development and implementation of three-dimensional computer-generated facial images capable of expression gesture and speech. A review of previous work in chapter one shows that to date the model of computer-generated faces has been one in which construction and animation were not separated and which therefore possessed only a limited expressive range. It is argued in chapter two that the physical description of the face cannot be seen as originating from a single generic mould. Chapter three therefore describes data acquisition techniques employed in the computer generation of free-form surfaces which are applicable to three-dimensional faces. Expressions are the result of the distortion of the surface of the skin by the complex interactions of bone, muscle and skin. Chapter four demonstrates with static images and short animation sequences in video that a muscle model process algorithm can simulate the primary characteristics of the facial muscles. Three-dimensional speech synchronization was the most complex problem to achieve effectively. Chapter five describes two successful approaches: the direct mapping of mouth shapes in two dimensions to the model in three dimensions, and geometric distortions of the mouth created by the contraction of specified muscle combinations. Chapter six describes the implementation of software for this research and argues the case for a parametric approach. Chapter seven is concerned with the control of facial articulations and discusses a more biological approach to these. Finally chapter eight draws conclusions from the present research and suggests further extensions.

Yes / Custody images are a standard feature of everyday Policing and are commonly used during investigative work to establish whether the perpetrator and the suspect are the same. The process of identification relies heavily on the quality of a custody image because a low-quality image may mask identifying features. With an increased demand for high quality facial images and the requirement to integrate biometrics and machine vision technology to the field of face identification, this research presents an innovative image capture and biometric recording system called the Halo. Halo is a pioneering system which (1) uses machine vision cameras to capture high quality facial images from 8 planes of view (including CCTV simulated), (2) uses high quality video technology to record identification parades and, (3) records biometric data from the face by using a Convolutional Neural Networks (CNN) based algorithm, which is a supervised machine learning technique. Results based on our preliminary experiments have concluded a 100% facial recognition rate for layer 34 within the VGG-Face model. These results are significant for the sector of forensic science, especially digital image capture and facial identification as they highlight the importance of image quality and demonstrates the complementing nature a robust machine learning algorithm has on an everyday Policing process.

Machine Learning (ML) is a research area in artificial intelligence (AI) and computer science. ML focuses on the use of data and algorithms to identify patterns in data without direct instruction. This is done with the help of ML algorithms that learn to make predictions by finding rules and drawing conclusions based on training data. ML can be used to perform tasks such as estimating human's age based on facial images, which can be used to control or restrict access to a website based on the user's age.Age estimation from facial images can be described as a regression problem or a classification problem. Estimating the exact age is a regression problem, while estimating the age group is a classification problem. A regression problem can be converted to a classification problem to determine the age group from the estimated age. This is done by dividing the total age range into different age groups, after which it is decided which group the age estimate belongs to. This study aims to answer how ML models can be used to estimate different age groups from facial images. This is done by exploring and evaluating two classification models that directly estimate the age group, in comparison with determining the age group from the exact age estimate by converting the regression problem into a classification problem. In this work, facial images are used to train and test ML algorithms by combining facial images from various open research databases. A delimitation was made in this study to only explore the use of Convolutional Neural Networks (CNN) to create different ML models that can estimate the age or the age group. CNN are used to perform tasks that require image interpretation, which in this case means that facial images are interpreted to make predictions. The results show that one of two classification models in this study achieves an accuracy of 75.9%. The second classification model, which estimates other age groups, achieves an accuracy of 62.88%. However, the outcome of two converted classification problems from a regression model shows an accuracy of 68.85% and 70.68%, respectively. The estimation model that achieves the highest accuracy when estimating the age group is a classification model with 75.90% accuracy. The work indicates that the choice of age group interval and facial images within each age group determine how the estimation models perform in relation to each other. / Machine Learning (ML) är ett forskningsområde inom artificiell intelligens (AI) och datavetenskap. ML fokuserar på användningen av data och algoritmer för att identifiera mönster i data utan direkt instruktion. Detta sker med hjälp av ML-algoritmer som lär sig att göra förutsägelser genom att hitta regler och dra slutsatser utifrån träningsdata. ML kan användas för att utföra uppgifter som att uppskatta människors ålder utifrån ansiktsbilder, vilket kan användas för att kontrollera eller begränsa åtkomsten till en webbplats baserat på användarens ålder. Åldersuppskattning från ansiktsbilder kan beskrivas som ett regressionsproblem eller ett klassificeringsproblem. Att uppskatta den exakta åldern är ett regressionsproblem, medan att uppskatta åldersgruppen är ett klassificeringsproblem. Ett regressionsproblem kan konverteras till ett klassificeringsproblem för att bestämma åldersgruppen från den uppskattade åldern. Detta utförs genom att dela upp det totala åldersintervallet i olika åldersgrupper, varefter det avgörs vilken grupp åldersuppskattningen tillhör. Denna studie ämnar svara på hur ML-modeller kan användas för att uppskatta olika åldersgrupper från ansiktsbilder. Detta sker genom att utforska och utvärdera två klassificeringsmodeller som direkt uppskattar åldersgruppen, i jämförelse med att bestämma åldersgruppen från den exakta åldersuppskattningen genom att konvertera regressionsproblemet till ett klassificeringsproblem. I detta arbete används ansiktsbilder för att träna och testa ML-algoritmer genom att kombinera ansiktsbilder från olika öppna forskningsdatabaser. En avgränsning gjordes i denna studie för att endast undersöka användningen av Convolutional Neural Networks (CNN) för att skapa olika ML-modeller som kan uppskatta åldern eller åldersgruppen. CNN används för att utföra uppgifter som kräver bildtolkning, vilket i det här fallet innebär att ansiktsbilder tolkas för att göra förutsägelser. Resultaten visar att en av två klassificeringsmodeller i denna studie uppnår en noggrannhet på 75,9%. Den andra klassificeringsmodellen, som uppskattar andra åldersgrupper, uppnår en noggrannhet på 62,88%. Däremot visar utfallet av två konverterade klassificeringsproblem från en regressionsmodell en noggrannhet på 68,85% respektive 70,68%. Den uppskattningsmodell som uppnår högsta noggrannhet vid uppskattning av åldersgruppen är en klassificeringsmodell med 75,90% noggrannhet. Arbetet tyder på att valet av åldergruppintervallet samt ansiktsbilder inom varje åldersgrupp avgör hur uppskattningsmodellerna presterar i förhållande till varandra.

Recent increase in widespread use of facial recognition technologies have accelerated the utilization of demographic information, as extracted from facial features, yet it is accompanied by ethical concerns. It is therefore crucial, for ethical reasons, to ensure that algorithms like face recognition algorithms employed in legal proceedings are equitable and thoroughly documented across diverse populations. Accurate classification of demographic traits are therefore essential for enabling a comprehensive understanding of other algorithms. This thesis explores how classical machine learning algorithms compare to deep-learning models in predicting sex, age and skin color, concluding that the more compute-heavy deep-learning models, where the best performing models achieved an MCC of 0.99, 0.48 and 0.85 for sex, age and skin color respectively, significantly outperform their classical machine learning counterparts which achieved an MCC of 0.57, 0.22 and 0.54 at best. Once establishing that the deep-learning models are superior, further methods such as semi-supervised learning, a multi-characteristic classifier, sex-specific age classifiers and using tightly cropped facial images instead of upper-body images were employed to try and improve the deep-learning results. Throughout all deep-learning experiments the state of the art vision transformer and convolutional neural network were compared. Whilst the different architectures performed remarkably alike, a slight edge was seen for the convolutional neural network. The results further show that using cropped facial images generally improve the model performance and that more specialized models achieve modest improvements as compared to their less specialized counterparts. Semi-supervised learning showed potential in slightly improving the models further. The predictive performances achieved in this thesis indicate that the deep-learning models can reliably predict demographic features close to, or surpassing, a human.