The Application of Evolutionary Algorithms to the Classification of Emotion from Facial Expressions

Emotions are an integral part of human daily life as they can influence behaviour. A reliable emotion detection system may help people in varied things, such as social contact, health care and gaming experience. Emotions can often be identified by facial expressions, but this can be difficult to achieve reliably as people are different and a person can mask or supress an expression. Instead of analysis on static image, the computing of the motion of an expression’s occurrence plays more important role for these reasons. The work described in this thesis considers an automated and objective approach to recognition of facial expressions using extracted optical flow, which may be a reliable alternative to human interpretation. The Farneback’s fast estimation has been used for the dense optical flow extraction. Evolutionary algorithms, inspired by Darwinian evolution, have been shown to perform well on complex,nonlinear datasets and are considered for the basis of this automated approach. Specifically, Cartesian Genetic Programming (CGP) is implemented, which can find computer programme that approaches user-defined tasks by the evolution of solutions, and modified to work as a classifier for the analysis of extracted flow data. Its performance compared with Support Vector Machine (SVM), which has been widely used in expression recognition problem, on a range of pre-recorded
facial expressions obtained from two separate databases (MMI and FG-NET). CGP was shown flexible to optimise in the experiments: the imbalanced data classification problem is sharply reduced by applying an Area under Curve (AUC) based fitness function. Results presented suggest that CGP is capable to achieve better performance than SVM. An automatic expression
recognition system has also been implemented based on the method described in the thesis. The future work is to propose investigation of an ensemble classifier implementing both CGP and SVM.