Facial Landmark Localization Under Large Pose Variation

Biometrics of face images has been one of the hot topics in the fields of artificial intelligence,machine vision and pattern recognition.It mainly consists of the detection of human face,the location of facial landmark,the gesture recognition of human faces,the analysis of facial expressions,Different composition between the faces,face recognition technology and live testing and face animation and other research components.In face recognition applications,face detection and facial landmark localization is the basis,but also the key to follow-up identification technology.The factors of affecting facial landmark localization are mainly external such as light exposure,occlusion and face different gestures and rich expressions,these factors which pose a great challenge to facial landmark localization,and is also the problem that face recgnition need urgent ly to solve in the actual project.In recent years,the deep learning in the image is widely used by researchers due to it has a strong characterization ability.Deep learning requires a large number of clean face images for training models.However,there are some noises in the open data set or the images captured on the Internet,which will greatly affect results of models.Based on the research of traditional clustering algorithm,this paper proposes a new algorithm based on large-scale face image deleting outlier samples,which can effectively denoise face training data and improve the accuracy of face feature extraction model in a certain extent.Secondly,this paper uses Gaussian mixture model to automatically align face images.The method mainly uses the deep features of face pictures to cluster the pictures,so that the face pictures of different head gestures and expressions can be clustered into different categories,and the pose of the training samples can be automatically classified.Then,the face feature matrix is constructed according to the different types of face images.In this paper,a mapping relationship is established between the face feature matrix and the corresponding predicted facial landmark,and the mapping matrix is calculated by the least squares method to predict facial landmark of inputing face images.