Face Completion And Facial Expression Recognition Based On Generative Adversarial Networks And Knowledge Distillation

The face plays an important role in the process of human interaction with the others,it conveys distinctive identity characteristics.Therefore,the research of face-related image algorithms is the most popular research direction in computer vision area.The face algorithms are also widely used in social field,entertainment,security,medical field and many other areas.With the development of the Internet and the large-scale applications of face-related image algorithms,the quality of face images is crucial,The application of damaged face image repair techniques has become more and more widespread.At the same time,facial expression recognition technology has gradually become a major subject in the field of human psychology and emotion research,and its effect in the research of human-computer interaction technology is increasing.In summary,face completion and facial expression recognition have played an increasingly important role in face-related image problems,with great influence in theory and practice.They have received widespread attention in the academic and industrial circles.The goal of this paper is to solve the problem of facial expression recognition with occlusion noise,which mainly consists of two parts: face completion and expression recognition.The classic generation adversarial networks for face completion cause the pixel value of the unbroken area to be shifted.In view of the above problems,this paper uses a noise segmentation method to optimize the classical generation adversarial network.Aiming at the problem of inaccurate noise segmentation,a method of knowledge distillation was proposed to optimize the noise segmentation network.As for the main emotion recognition deviations of facial expressions by classical methods,a knowledge distillation method based on pseudo-twin network was proposed for optimization.In addition,due to the lack of a public data set for joint face completion and face expression recognition tasks,this paper makes an occlusion mask based on the existing expression recognition dataset and constructs a face data set containing occlusion noise.the main work is divided into the following parts:(1)Optimize face completion method based on classic generative adversarial network.Aiming at the limitations of the classical generative adversarial network,an occlusion noise localization method was proposed,and the fixed-point repair method was used to alleviate the pixel value shift of the unbroken area after the original image was filtered several times.Experiments show that this method can further improve the accuracy of face completion.(2)Optimize the face noise segmentation model U-Net++.The teacher-student distillation network is constructed based on the sub-models with different parameter quantities in the original U-Net++.The teacher model is used to provide soft label for the student model and enrichs the supervision information.Experiments show that this method can improve the performance of U-Net++ sub-models.(3)A knowledge distillation method based on pseudo-twin networks.In order to analyze the main tendency of sample emotions,an end-to-end pseudo-twin distillation network is designed and used in the facial expression recognition task.(4)Based on the existing facial expression database,create 15,339 samples with face occlusion noise to form a new dataset.