The Facial Feature Point Estimate Algorithm Based On Regression Research

Face annotation is one of the important processes in the application of face recognition. Face annotation aims to estimate the spatial locations for face images under varying poses, expressions and illuminations. Meanwhile, the rate of the face recognition is very sensitive to the accuracy of the estimation for facial landmark points. Hence, automatic face annotation is a challenging research problem.This dissertation studies regression-based estimation algorithms for facial landmark points, especially focuses on estimation algorithms for facial landmark points under various poses and expressions. This dissertation estimates the non-frontal or non-neutral facial landmark points based on the frontal or neutral ones using the different regression models.The main contributions of this dissertation include:1) We propose the linear regression models with neighboring samples to make full use of the neighboring information between the testing samples and the training samples. The linear regression models with neighboring samples include the dense linear regression model with the neighboring samples, the sparse linear regression model with the neighboring samples and the fully sparse linear regression model with the neighboring samples. The experimental results show that the estimation algorithm based on the linear regression models with the neighboring samples outperforms the estimation algorithm based on the existing linear regression models.2) We propose the patch-based regression model to estimate the facial landmark points based on the differences of the facial features in different facial regions and the similarities of the facial features in the same area of the human face. The experimental results show that the performance of the proposed model is better than that of the dense linear regression model, the fully sparse linear regression model, the dense nonlinear regression model and the fully sparse nonlinear regression model.3) We propose the neighborhood-preserving estimation algorithm for facial landmark points because the neighboring structure of the facial shapes is invariant to pose changing. The neighborhood-preserving estimation algorithm for facial landmark points is based on the following assumption:the neighboring structure of the face shapes in non-frontal view is consistent with that in frontal view. A face shape in frontal or non-frontal view can be represented as a linear combination of its neighbors. The experimental results show the superiority of the neighborhood-preserving estimation algorithm for facial landmark points to the algorithms based on the existing linear regression models and the existing nonlinear regression models when estimating the facial landmark points under varying poses.4) By combining regression-based estimation methods for facial landmark points with the texture mapping, we generate the virtual face images at various poses and expressions. The experimental results show the good generality of the regression-based estimation algorithms.