Face Alignment and Face Muscle Movement Identification Using GAN

Ng, Han Xiang (2022) Face Alignment and Face Muscle Movement Identification Using GAN. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Ng Han Xiang.pdf Restricted to Registered users only Download (3MB)

Abstract

The conventional face recognition algorithms are unable to perform well under pose variations inputs. The locations and the visibility of the key facial features such as eyes, nose, mouth and etc. will be affected under extreme head pose, illumination, expression and etc. causing the conventional face recognition network failed to recognize the identity. Therefore, pose-invariant face recognition (PIFR) has become a major challenge for conventional face recognition algorithms. PIFR is relatively important for video surveillance such as watching people in Home-nursing, shopping malls, education institutions, hospitals, waiting rooms and monitoring vehicles. This is because, in most of the scenarios, free-walking people would not always keep their faces frontal to the cameras. The topic of Generative Adversarial Network (GAN) have received more attention lately because GAN is one of the techniques that is widely used for face frontalization in order to perform PIFR. In this thesis, a TP-GAN is proposed to perform head pose correction of the given profile images. The network structure design for the TP-GAN consists of two pathways which is the local and global pathway. The local pathway is used to preserve the local facial features such as the eyes, nose and mouth while the global pathway is used to synthesize the global structure of the face. The training results of the TP-GAN are considered acceptable as the model able to generate the frontal view of the individual as the ground truth images. The testing results of the TP-GAN has room to be improved as the generated images are having huge difference with the ground truth image. The symmetrical loss exhibits unexpected increase during the training of the TP-GAN.