Research On Pose And Expression Invariant3D Face Recognition

Automatic face recognition technology is widely applied in many applications, such as identity authentication, access control and financial security. After decades of research, face recognition on2D images has developed a lot, but still suffers from a lot of challenges, such as illumination, pose and expression variations. Comparing to2D images,3D images aren’t projected in the process of acquisition so they are easier to be aligned, and don’t include intensity information so they are free from illumination. Recently, with the rapid advance of3D image acquisition technology, many researchers have turned to face recognition on3D faces, in the hope of overcoming the challenges of2D recognition.Our research is focused on pose and expression invariant3D face recognition. Expression variation deforms the3D face, and transforms the facial geometric shape; in the condition of pose variation, the facial data will transform according to the rotating angle, and some parts of the face will lose big amount of data because of self-occlusion. Both the two conditions will enlarge the intra-class distance of two faces from the same identity and the intra-class distance may become larger than interclass distance, which will decrease the face recognition accuracy. So we propose a coarse-to-fine face registration method to align the faces to be neutral, and adopt feature extraction method based on rigid part of face to overcome expression variation. The main research work and innovative contributions are as follows:1. We have constructed average nose models based on depth images to detect noses. First, we segment the facial parts through Random sample consensus algorithm, preprocess them to improve quality and transform them to depth images, then extract the most rigid part of faces-the nose part to construct three average nose models, not only to detect the noses of faces with pose, but also to calculate the rough rotation angles, and then coarsely align the faces.2. We have constructed an active appearance model based on depth images to finely align the faces. Based on facial shape and appearance of depth images, we construct an active appearance model based on depth, best match the faces after rough alignment with the model using an iterative method, and finely align the faces.3. We have proposed a depth-based direct feature extraction method and also used LBP feature extraction method. To decrease the expression variations on features, we simply extract the features on relatively rigid upper face, and to handle loss of facial data, we apply multiple regions fusing method. Finally we train the features by support vector machine method, obtain the classification model and complete face recognition.