Discriminative Sparse Representation for Expression Recognition

This thesis is focused on recognising emotions of different subjects through facial expressions in 2D images. We will go through the multiple stages of this problem where we aim to take maximum advantage of supervised algorithms and labelled information. We will compare different pixel processing techniques and show that the histogram based ones, like HOG and LBP, have the best performance for this particular problem. Sparse representation has definitely been proved to be a very good way to solve computer vision problems in facial understanding over the last couple of years. Therefore we will make use of a new label consistent singular value decomposition algorithm to learn a discriminative dictionary and compare its performance with several supervised dimensionality reduction techniques. Finally we will obtain state-of-the-art classification accuracies for the problem of recognising facial expressions with our histogram supervised manifold preserving sparse representation technique. We will test different methods across multiple databases containing images of various subjects performing various expressions, aligned or non-aligned.