Lip segmentation with hybrid model

Image segmentation is a principal research area in computer vision. Several methods have been expanded in segmentation fields and these can be classified into three categories which are active appearance models, active contour models, and active shape models. In this paper, we are interested in lip segmentation. It is an essential step for many multimedia systems such as lip reading, low bit rate communication systems or video conferencing. We will present an automatic algorithm concept for lip segmentation used for video sequences. Our approach is a hybrid model of active contour based on the geodesic active contour (GAC) and a distance regularized level set evolution (DRLSE) model. The hybrid model is a variational level set approach that utilizes the edge information to find lip contours by minimizing the energy function. The hybrid model uses a Gaussian filter to denoise the image and this will produce fuzzy image edges. In addition, the model still can't segment properly. To solve these problems, we will design a filter from the combination of median and average filters, which can reduce the noise while preserving edge information. In our proposed approach, we localize the face and detect a bounding box of the mouth succeeded by applying our hybrid model to extract the key points of mouth. As a post-treatment, we apply a parametric model to enhance the full contour of the lips. The parametric model is based on the cubic curves to determine the possible deformations of the lips. The experiments show that our algorithm is effective and practical. The proposed approach has high segmentation efficiency compared to the classic hybrid model with a Gaussian filter.