Research On Precise Facial Landmark Localization Based On Optimal SIFT-based Feature Discriptors

Face image analysis is a frontier topic in the field of computer vision and pattern recognition,including a series of researching content,for example,face detection,facial landmark localization,facial expression recognition,face recognition and so on.Among them,face detection and accurate facial landmark localization are the foundation and important premise of applications such as face recognition,expression recognition and attitude estimation.They are the key problems needing resolving in the process to expand the scope of applications based on the analysis of the human face as well.Considering the practical demand,our research is to detect and locate face,segment the oval face region from the background,and then do the accurate facial landmark localization basing on special textural characteristics and shape constraints over the oval face region.Traditional facial landmark localization algorithm haven't studied too much about the specific textural characteristics of facial landmarks,so the initial detection of the facial landmark is not good,which directly affects the subsequent shape constraints and makes it difficult to improve the accuracy of facial landmark localization.Considering the problem mentioned above,firstly,this paper adopted the Powell algorithm to obtain the parameters that optimized the Scale Invariant Feature Transform(SIFT)-based landmark-specific feature descriptors in accordance with their special textural characteristics.Secondly,we trained support vector machine(SVM)regressors with the landmarks' SIFT-based feature descriptors calculated based on those learned parameters.Thirdly,we detected facial landmarks by the regressors over the ROI.The paper optimised the parameters of the Scale Invariant Feature Transform(SIFT)-based landmark-specific feature descriptors with LFPW database as the training set.And then we trained support vector machine(SVM)regressors with the landmarks' SIFT-based feature descriptors calculated based on those learned parameters.Finally we tested the regressors on the Bio ID database.It is proved that,the initial localization of the facial landmarks performed by the support vector machine(SVM)regressors trained on the optimized landmarks' SIFT-based feature descriptors is much more accurate than the support vector machine(SVM)regressors trained on the original landmarks' SIFT-based feature descriptors.Besides,we learned triplet-based shape model by Bio ID database to correct potential outliers of the initial localization of the facial landmarks,making the facial landmark localization more accurate.