Research Of Face Image Inpainting Based On Gan Method

Image inpainting is a common image editing operation,which aims to fill the missing or masked areas in images with newly generated contents.The newly generated contents can be as accurate as the original images,and can also be simply adapted to the surrounding pixels,so that the completed image looks real.Face inpainting is a challenging task,especially for the case of filling large missing regions or essential components of faces such as eyes,noses,mouths and so on.In recent years,the image inpainting method based on GANs has achieved remarkable performance.The Generative Adversarial Networks framework can not only generate new pixels that have never appeared before,but also generate clear and realistic images.In addition,it has a good performance for repairing the large area missing images or the missing area of arbitrarily shapes.In this paper,a new face image inpainting algorithm based on Generative Adversarial Networks is designed to solve the problem of large missing face images or important parts of missing face images.Our paper is based on the Deep Convolution Generative Adversarial Networks(DCGANs),and the loss function and network are improved to improve the effect of image inpainting.In this paper,Wasserstein distance and Lipschitz constraint ——gradient penalty,are introduced to improve the loss function,which raised the stability of model training and solved the problem of mode collapse during network framework training.Besides,in the discriminator,Layer Normalization is adopted to replace Batch Normalization to avoid the mutual dependence of different samples in the same batch during training.In order to obtain more accurate high-level representation and generate shaper outline contents,this paper also adds a local discriminator on the basis of the global discriminator.After the image is generated,the newly generated image and the damaged image are fused.For the purpose of making the fusion boundaries look less obvious and the whole image looks more real,another poisson fusion algorithm which called PIE with better effect is also cited in this paper.In this paper,CelebA face data set is used as training set to train the netw ork framework,and LFW face data set is used as prediction set to build damaged images after mask processing and be input into the model for repair.In the contrast experiment,qualitative and quantitative analysis was made from four different damaged areas.Multiple experimental results show that the improved image inpainting algorithm can effectively improve the quality of restored images.The PSNR values of the improved model in the center damage area,left damage area,left eye damage area and mouth damage area have reached 24.3 db,23.3 db,28.7 db and 25.6 db respectively.SSIM values have reached 0.781,0.761,0.952 and 0.828 respectively.