Semantic Image Inpainting Based on Autoencoding Adversarial Networks

Image inpainting is one of the most challengeable tasks in computer vision study. Methods based on image statics of unmasked area generally produce smooth results, but has no semantics understanding of the images. Methods based on deep learning technic have semantics understanding of images, but obtain results with blurred inpainting content. We propose an inpainting network based on autoencoding networks to repair masked images semantically with a combination of encoder-decoder network and mix-discriminators in this study. The mix-discriminators are composed of a global and a local discriminator to judge whether the results are completed or not. In detail, we train an encoder network to learn and extract hidden representations from the input images, and a decoder network to predict the contents of the masked areas based on the hidden representations. Then the outputs and the original images are fed to a global discriminator for judging the consistency of whole output images and a local discriminator for supervising the authenticity of the inpainting areas. The two discriminators will propaganda back a gradient to train the inpainting network to produce realistic image contents.