Face Inpainting Based On Residual-wasserstein Generative Adversarial Networks

Digital face repairing refers to the process of recovering defective facial images using computer means.According to certain rules,the image is repaired with the aid of the pixel information of the defective area,so that the observer cannot distinguish whether it is a corrupted image.Image repairing has applications in multiple directions,such as:repairing of old images,removal of obstructions,repairing of video information,etc.The images restored by traditional image repairing techniques often have problems such as blurring and artifacts;when the defect area is large,the repair work cannot be completed well,too.When it comes to the repairing of face images,it is no longer possible to use the principle of information diffusion to predict corrupted areas.This paper proposes a face image repairing algorithm based on Residual-Wasserstein generation adversarial network.The innovation of this algorithm lies in its ability to obtain background information from deep networks unsupervised to complete image restoration.When the defects such as the blurring of the traditional repair algorithm are improved,the effect of image generation and repair is further improved by using deep learning methods.The main work of this article is as follows:(1)Analyze and study the training methods and model structure of traditional generative adversarial networks.Although the network has excellent sample generation effects in unsupervised learning,it also has many defects,including gradient dispersion,unstable training process,and difficulty in convergence problem.The DCGAN introduced in this paper can use generative adversarial networks for image processing,and WGAN can improve the problems of difficulty in convergence and collapse in training.This paper improves and trains the traditional generative adversarial network based on these two networks.(2)Simultaneous introduction of residual network and spectral normalization algorithm.Using the residual network to deepen the number of learning layers without degrading the training,this paper proposes the Residual-WGAN algorithm to construct the training network.Secondly,the spectral normalization algorithm is introduced to make the network satisfy Lipschitz continuity without changing the structure of the parameter matrix,and enhance the training ability.(3)This paper defines a trinity loss function to train the network by context loss,perception loss and Wasserstein loss.First,randomly select 50,000 data into the network for training in Celeb A,and then input the image to be repaired into the trained network to generate a series of fake images Then,the best fake image is selected from the context and perceptual losses.Therefore,the image is repaired by patching into the corrupted image.Finally,the repairing effect is verified by PSNR and SSIM methods.The PSNR value of the repaired image is 22.23dB,and the corresponding SSIM value is 0.8609.Compared with the traditional DCGAN network repair effect,the PSNR and SSIM values are increased by 2.86dB and 0.0528,respectively.You can also get clearer visual effects.