Research On Eyeglasses Removal And Face Restoration Based On Adversarial Networks

After years of research,face recognition technology has reached a level close to that of the human eye,and glasses as a common accessory on face images will easily interfere with the processing of images.In such cases,the effects of face recognition are often affected.In particular,criminals often use glasses to camouflage which seriously affects the matching of their identities.In recent years,as the emergence of generative adversarial networks,image translations has gradually become one of the focuses of researchers.For this reason,this thesis uses image translation as to solve the problem of the removal of glasses on face images.By studying existed unsupervised image generation models,we propose an image translation model of eyeglasses removal uses the unsupervised sample learning methods and can retain the facial features to the maximum extent.First,with the paired faces samples with/without glasses difficult to obtain,we use neural networks to obtain samples with/without glasses from the face samples in the CelebA library to establish the training dataset.The trained classification model is used as an objective evaluation criterion for evaluating the effect of the eyeglasses removal.Moreover,thousands of paired samples with/without glasses were collected in this thesis,and a subjective scoring system is established for evaluating the eyeglasses removal effectiveness.Secondly,after compared different generational networks,we discover that the generator of the cycle generative adversarial network(CycleGAN)has strong learning ability in image distribution by combining sample instance normalization and residual learning modules.This thesis selects CycleGAN as the basic experimental network,and a skip connect is configured between the encoding layers and corresponding decoding layers in the generator network,so that the underlying information of the image can be shared in the whole network,and the feature information of the input image can be retained.Experiments results show that the addition of the skip connection model improves the performance of glasses removal and keeps the facial feature information.Finally,in order to better ensure that the network can concentrate on learning the distribution of face glasses regions under unpaired samples,this thesis introduces a distance loss as a constraint in the training process,which seeks to greatly retain the feature information of other areas of the face during the process of the glasses removal.To get ideal results,this thesis changes the calculation of the distance with the distance between the upper and lower parts of a single sample instead of the original.The experiments results show that in both the subjective and objective criteria,our trained model can effectively remove the glasses from the face image and is better than other methods in preserving the original features of the human face.