Large Pose Detection and Facial Landmark Description for Pose-invariant Face Recognition

Face recognition is an important computer vision task affected by several factors. These factors include the face pose of the input image, features used to describe the image, illumination conditions, and facial expression. In this study, a pose-invariant face recognition framework based on large pose detection and facial landmark description is proposed. During the training phase, a large pose detector model is proposed to process the 2D spatial distributions of the detected facial landmarks on a set of face images. This model can detect whether the yaw angle of the face is large or small (semi-frontal face image). This results in two face pose scenarios. Then, a feature descriptor is applied to a set of predefined facial landmarks on a face image for obtaining the feature vectors. These feature vectors are used to train two face recognition models for each person in the database. One for the large pose scenario and the other for the semi-frontal pose scenario. During the testing phase, the large pose detector is used to select a type of face recognition model (large pose or semi-frontal one). The selected model is utilized to determine the identity of the person. In this study, the CMU-PIE database is employed. Three feature descriptors, SIFT, HOG, and LBP, are adopted for comparison. The models used for face recognition are SVM, GMM, and Naive Bayes. The novelty of the proposed method is using a large pose detector to improve the face recognition rate. After performing experimental trials on face images with pose angles ±90°, a performance comparable with state-of-the-art methods is obtained.