SemanticStyleGAN: Generative Image Inpainting Using Style-Based Generator

Abstract

AbstractImage inpainting, a task of reconstructing missing or corrupted image regions, has a great potential to advance the fields of image editing and computational photography. Despite being a difficult problem, the researchers have made headway into the field thanks to the progress in representation learning with very deep convolutional neural networks and the successes in generation of realistic images by using Generative Adversarial Networks (GANs). To avoid the problems of generating blurry and distorted regions as well adding artefacts, a new model for image inpainting is proposed in this paper. The GAN-based architecture incorporates two frameworks, the model for semantic image inpainting and the models for style transfer, which have been pointed out to be successful in previous research. Our model was evaluated on the Flickr-Faces-HQ dataset. The results are promising and point out that using a combination of various GAN-based technologies could improve the performance on the task of image inpainting. Directions for future research are also discussed.KeywordsImage inpaintingGenerative adversarial networksStyle transfer