Eyeglasses Transfer Based On Genera-Tive Adversarial Networks

Occlusion face recognition is an important problem in face recognition.The past researches mainly focus on two aspects:classifier design and feature extraction.This paper attempts to ex-plore from the perspective of data augmentation.Taking the occlusion of glasses as an example,this paper studies the face attribute transfer by using the generative adversarial networks.A large number of images with glasses are added to supplement the training set,which reduces the inter-ference of glasses on face recognition and improves the accuracy of recognition.The main contributions of this work are as follows:.Block structure is used to replace the convolutional layer in the reference network.So neu-ral networks is deeper while computational cost is equal,and better quality images can be generated..It is proposed to add reconstruction constraints for non-eye region,so as to suppress wrong extraction and separation of non-eye region's features..This paper proposes a scheme to filter source input images,source images used for exchange and generated images through classifier and discriminator..Transfer learning method is used to train neural network in the form of two steps.Finally,a large number of photorealistic images can be generated to supplement the training dataset.