Dermatologically Inspired Deep Face Analytics

Deep Learning algorithms are a subclass of machine learning which uses large neural networks to make predictions from data, and even generate new data. Owing thanks to recent advancements in the field, deep learning is being applied through the physical sciences.

The aims and objectives of this project are interdisciplinary at their core. Ranging from novel contributions in data annotation, to neural network architecture; to dermatological analyses. We intend to explore the active research area of ageing and vitality, covering gaps in existing research and propose a new deep learning model to analyse the heterogeneous ageing of human faces. While there is existing work detailing both human and computational methods for grading face ageing, no research has been done to apply deep learning to this problem.

The ability to detect face ageing which is localised to specific region of the face has tangible benefits throughout both the medical and research communities. In medical practice it could be used to inform the diagnosis of conditions, and subsequently the speed and accuracy of screening. In the world of research, it allows analysis to be carried efficiently out over vast datasets, opening doors for new dermatological research which would have previously been cost prohibitive.

The main challenge in training any machine learning problem is data. In order to train a deep neural network to predict something accurately, there must be a large number of known data points. When approach the field of ageing, there is no dataset in existence labelling the rate at which difference face regions have aged. For this reason no attempts have yet been made to train a model such as the one we propose.

Our approach to this lack of data is to build on the work of others who have developed models for global face ageing, and for semantic face segmentation. Semantic face segmentation is a field which splits face regions at a pixel level. We intend to learn from both existing datasets of segmented faces and datasets containing global face ages. The combination of these two types of data provide a framework to develop novel new architectures for grading the ageing of faces at a feature level.

Another question we intend to answer during the course of this research is if the use of much higher resolution face images improves the quality of age prediction models, due to increased detail like wrinkles and spots. To do this we propose a novel experiment to efficiently annotate a new dataset with over 4x the resolution. We use the human ability to compare the age of faces to create annotations, rather than requiring that annotators are able to absolutely predict the age of a face.

Finally we validate the reliability of our system by testing it on many different images of the same individual, at the same age. Giving an uncertainty value that represents the viability of the tool in different environments.
A secondary, more novel approach to validating our model is using image manipulation to swap sections of faces between individuals, such as adding a 60 year old eyes to a 30 year old face. Using age estimation models we will be able to detect subtle changes in the global age of the face, these observations can be attributed to the segment which was swapped.

In summary, we aim to discover if deep learning is a valid replacement for existing methods in dermatological analysis. We aim to create a model for regional skin age grading, evaluating a range of different architectures and methods for pixel level classification. We propose a new method for annotating originally ranked data, optimising the human element by minimising the cognitive load.