A Method of Semantic Image Inpainting with Generative Adversarial Networks

Abstract

Semantic image inpainting focuses on the completing task of high-level missing regions at the basis of the uncorrupted image. The classical methods of image inpainting can only deal with low-level or mid-level missing regions due to the lack of representation of the image. In the essay, we conclude a new method of semantic image inpainting. It’s based on the generative model with learning the representation of image database. We propose an architecture of completion model using perceptual loss and contextual loss based on generative adversarial networks after having trained generative model using DCGAN. We qualitatively and quantitatively explore the effect of missing regions of different types and sizes on image inpainting. Our method successfully completes inpainting tasks in large missing regions and results looks realistic with extensive experiments. We conclude that the performance of our model mostly is good when completing image corrupted with the mask with an area of less than 50% as well as with center or random masks.