Face Adversarial Restoration For Facial Expression Analysis In Wild Environment

Nowadays the researches on facial expression are mainly direct analysis of static face images.These methods often perform well on the face image dataset in the laboratory environment,but it is difficult to apply these methods to the wild face dataset with damaged face,such as low resolution and partially occluded face images.This paper aims to restore the damaged face through the face restoration model,and then uses the facial expression recognition model to detect facial action units of the restored faces.The main problem solved in this paper is to ensure that the facial expressions of the restored faces are same as the real facial expressions,so as to ensure that the accuracy of facial expression recognition on the damaged faces in the wild environment is high.For the face restoration model,this paper mainly designs two methods based on the generative adversarial network:One method is to add the priori facial action units information to the face restoration model as conditional information through weak supervision;Another method is to restore the entire damaged face image by learning the damaged area from the undamaged area through the improved graph convolutional network.This design is mainly based on relations of coexistence or mutually exclusive between facial action units.For the facial expression recognition model,this paper uses the end-to-end recognition model to analyze facial expressions directly,and uses it to perform facial expression recognition on the restored face to supervise the facial actions of the generated faces.This paper executes experiments on the two public datasets:BP4D and DISFA.The effectiveness of the proposed face restoration model is verified qualitatively and quantitatively.