Facial expressions convey a wealth of information about our feelings, personality and mental state. In this thesis we seek efficient ways of representing and analysing facial expressions of varying intensities. Firstly, we analyse state-of-the-art systems by decomposing them into their fundamental components, in an effort to understand what are the useful practices common to successful systems. Secondly, we address the problem of sequence registration, which emerged as an open issue in our analysis. The encoding of the (non-rigid) motions generated by facial expressions is facilitated when the rigid motions caused by irrelevant factors, such as camera movement, are eliminated. We propose a sequence registration framework that is based on pre-trained regressors of Gabor motion energy. Comprehensive experiments show that the proposed method achieves very high registration accuracy even under difficult illumination variations. Finally, we propose an unsupervised representation learning framework for encoding the spatio-temporal evolution of facial expressions. The proposed framework is inspired by the Facial Action Coding System (FACS), which predates computer-based analysis. FACS encodes an expression in terms of localised facial movements and assigns an intensity score for each movement. The framework we propose mimics those two properties of FACS. Specifically, we propose to learn from data a linear transformation that approximates the facial expression variation in a sequence as a weighted sum of localised basis functions, where the weight of each basis function relates to movement intensity. We show that the proposed framework provides a plausible description of facial expressions, and leads to state-of-the-art performance in recognising expressions across intensities; from fully blown expressions to micro-expressions.
|Queen Mary, University of London
|Electronic Thesis or Dissertation
Page generated in 0.0021 seconds