ERGAN: High Perform GAN for Eyeglasses Removal

To solve the occlusion of eyeglasses on the eye area, which brings great challenges to face recognition and gaze tracking. Based on the successful application of face attribute editing task in Generative Adversarial Network, this paper proposes a new eyeglasses removal method ERGAN. The latent space of GANs has rich semantic information. By manipulating latent codes, different attributes of images can be changed. The key is to find the appropriate semantic direction. First of all, in order to enable GAN to process real images, we introduce the more advanced GAN inversion model IDInvert, which can convert the given real image back to the latent space of the pre-trained GAN model. Then use our model ERGAN to remove eyeglasses attributes. Subsequently, to find the appropriate semantic direction of eyeglasses editing, we add the semantic direction of the eyeglasses instance to the semantic direction calculated by InterfaceGAN. Recombine the two semantic directions into the semantic direction of eyeglasses editing. Finally, experiments are performed on GANs generated images and real images. The experimental results show that our method is helpful to improve the accuracy of eyeglasses removal.