Human face analysis

Abstract

This thesis presents several original authors' contributions related to two topics of human face analysis, namely face detection task and facial expression classification task, respectively. The original work is presented as two distinct parts. In the first part of the thesis, a method for improving the accuracy of Support Vector Machines for face detection is introduced, foilowed by a rigorous statistical analysis of its stability in the attempt of using the bagging approach for gaining superior classification performance. The second and the biggest part of the thesis are dedicated to the feature extraction topic appiied for facial expression recognition. Independent component analysis is a tool used in this regard. Several linear and non-linear independent component analysis methods are investigated and compared, and interesting conclusions are drawn. Next, two novei non-negative matrix factorization algorithms are described and their ability for providing useful features for classifying facial expression is proven through extensive experiments. By analogy to neurophysiology, the basis images discovered by non-negative matrix decomposition couid be associated with the receptive fieids of neuronal cells involved in encoding human faces. Taken from this point of view, an analysis of these three representations in connection to the receptive fieid parameters such as spaţial frequency, frequency orientation, position, length, width, aspect ratio, etc, is undertaken. By analyzing the tiling properties of these bases some conclusions of how suitable these algorithms are to resemble biological visual perception systems can be drawn. The thesis ends up with a new feature extraction method using the phase congruency concept for measuring the similarity between image points, also appiied for facial expression recognition.