Facial Feature Points Localization Algorithm Based On Support Vector Regression

Facial feature point detection is a key step in face recognition, the accuracy of facial feature points directly affect the accuracy of face recognition. Facial features localization is a challenging research, because many factors, such as changes in illumination, different expressions, gestures, etc. will affect the accuracy of the facial feature points localization. The research paper focuses on the study of feature points localization. The main contents are as follows:(1) We use Support Vector Regression(SVR) algorithm for facial features localization. The method needs a trained model.First,we get one vector displacement of the test point and the manually target point.Then we construct the model by training which is the relationship between the displacement and the texture features of the test point. According to the trained model,we can predict the target point by the texture features of the test point. One of the major drawbacks of SVR is that it does not provide a probability of the estimate obtained which will affect the accuracy of the algorithm. To tackle the problem,we use an evaluation function to evaluate the probability of the estimated point to be the target location. So we can give up the point with low probability to improve the accuracy of the algorithm.(2) We combines Gabor wavelet transform and Local Binary Pattern(LBP) to describe the texture feature of the test point. First we do Gabor wavelet transform in 2scales and 4 directions to the 18×18 region centered on the test point to get Gabor feature images. Combine the Gabor feature images under two scales and then the Gabor feature images are segmented to 3×3 parts and encoded by Local Binary Pattern in two scales respectively.We Statistics the LBP feature of every segmented part to get one LBP histogram. At Last,we connect all of the histograms in a certain order to describe the texture feature of the test point. Given the feature dimension of our method being large, we use Correlation based Feature Selection to remove redundant or irrelevant features.(3) We propose the information updated sampling strategy based on the estimated point. during the estimation, we will evaluate the estimated point produced by SVR at each iteration.If the assessment value of the point is high,we will add the point to the information set to update the sampling region. The updated sampling cansample from the region close enough to the target location which improves accuracy of SVR.So this strategy can effectively improve the positioning accuracy.(4) Finally, our algorithm is tested in different databases and the experiments results show the feasibility and effectiveness of the algorithm. and our algorithm have good robustness about the variation of expressions, illumination, and posture.