Statistical facial feature extraction and lip segmentation

Official URL: http://192.168.1.20/record=b1293776 (Table of Contents)

Facial features such as lip corners, eye corners and nose tip are critical points in a human face. Robust extraction of such facial feature locations is an important problem which is used in a wide range of applications including audio-visual speech recognition, human-computer interaction, emotion recognition, fatigue detection and gesture recognition. In this thesis, we develop a probabilistic method for facial feature extraction. This technique is able to automatically learn location and texture information of facial features from a training set. Facial feature locations are extracted from face regions using joint distributions of locations and textures represented with mixtures of Gaussians. This formulation results in a maximum likelihood (ML) optimization problem which can be solved using either a gradient ascent or Newton type algorithm. Extracted lip corner locations are then used to initialize a lip segmentation algorithm to extract the lip contours. We develop a level-set based method that utilizes adaptive color distributions and shape priors for lip segmentation. More precisely, an implicit curve representation which learns the color information of lip and non-lip points from a training set is employed. The model can adapt itself to the image of interest using a coarse elliptical region. Extracted lip contour provides detailed information about the lip shape. Both methods are tested using different databases for facial feature extraction and lip segmentation. It is shown that the proposed methods achieve better results compared to conventional methods. Our facial feature extraction method outperforms the active appearance models in terms of pixel errors, while our lip segmentation method outperforms region based level-set curve evolutions in terms of precision and recall results.