3D Morphable Model-based Multi-view Face Recognition And Its Application

In this era,face recognition technology has been widely used and relatively mature in some limited scenarios.However,there are still some defects in this technology.The main manifestation is that most of the current face recognition products have high requirements for posture,users are required to actively adjust their posture to the front in order to obtain better recognition performance,which greatly limits the promotion of this technology in more applications.Therefore,multi-view face recognition technology is of great significance in many non-user cooperation application scenarios.Regarding the above problems,this thesis carried out the multi-view face recognition,studied and implemented a face rotation technology based on threedimensional morphable model,which can reduce the pose requirements of face recognition by frontal face alignment of side face images with large pose.The main work of this thesis is summarized as follows:1)Based on the lightweight networks and the additional angular margin loss function,the face recognition model training was carried out,and the influence of network capacity,face posture changes and other factors on the model accuracy was explored by model evaluation experiment,where a lightweight face recognition model having high performance was implemented.2)A face rotation technology that is based on 3D face reconstruction was studied and implemented to improve the problem that face recognition is limited by pose change.A Rotate and Render framework was used to generate paired face data for training generative adversarial network to realize the self-supervising of single face image,which solved the problem that it is difficult to obtain paired face data from multiple views,and at the same time avoided the problem of over fitting caused by using specific datasets for training.A generative adversarial network was designed and trained,which can generate realistic face images with better retention of identity information.Compared with other similar methods,this method achieved better identity information retention performance.Based on the fusion distance measurement,the effectiveness of the method was verified.3)Applied face rotation technology to complete the face alignment process of side face image in a real scene,designed and implemented a face recognition system for blind people in non-user cooperation scene,which reduced the requirements for posture.