Face Inpainting based on Improved WGAN-modified

Image Inpainting aims to use the technical methods to repair and reconstruct the corrupted region of the corrupted image, so that the reconstructed image looks more authentic. In this paper, the improved Wasserstein Generative Adversarial Network combined with the perceived loss and context loss function is used to repair the corrupted region of the image. By finding the nearest neighbor coding method from the defective image in the latent space and using the improved WGAN to add the spectral normalization constraint, the two main problems of the original GAN including the pattern collapse and gradient dispersion can be greatly reduced, in which can produce more clear as well as realistic images. Test experiments on the CelebA face data set indicate that the method has a good inpaint result.