Research On Face Restoration Method Based On Face Attributes

In recent years,people need to wear masks.However,at this stage,face recognition networks,especially with the impact of COVID-19 on many occasions.However,current face recognition networks can only recognize complete face images,and cannot get the identity information from masked face images.To make current face recognition networks work normally,it is necessary to translate masked face images to normal face images.At present,face restoration networks are usually used to realize this function.This research can not only enable the current face recognition networks to recognize the masked face images but also can be used on many occasions such as video shooting,which has very important practical application and theoretical research value.Face restoration network technology generally adopts a deep-learning method to extract effective feature information from face images,and generate similar images according to the feature information.The research of face unmasking using a face restoration network is a relative technology.Due to the difficulties of detecting various kinds of masks and face images that have complex semantic information,the deep learning method used in face restoration network technology at this stage can not unmask masked face images well.In view of the above difficulties,based on the deep learning method of the generative adversarial network,we proposed a face restoration network to remove the mask in the masked face images and convert them into normal face images.The main research is described as follows:1.In order to solve the problem that there is a difference between the mask area and the rest area in the unmasked face image using the current face unmasking method,we designed a face restoration network consisting of two subnetworks: the face segmentation subnetwork and the face completion subnetwork.By adding the face mask region detail loss function and face attribute loss function,our network effectively reduces the gap in attributes and details between the mask area and the rest area of the unmasked face image,which successfully solves the problem of complex face semantic information.2.As there are a variety of masks,which makes the current face restoration network cannot completely segment different types of masks,we innovatively design a mask translation branch that replaces the mask of the masked face images in the testing dataset with the mask in the training dataset,which successfully solves the problem of mask remaining in unmasked face images.The unmasking results on the simulated masked face testing dataset and real-world masked face testing dataset demonstrate that our face restoration network outperforms numerous state-of-the-art methods in highquality face recovery with identity preservation,has very high generalization,and can meet the requirements of face recognition.