Semantic image inpainting based on Generative Adversarial Networks

Abstract

Semantic image inpainting is different from traditional methods. Although traditional image inpainting methods can achieve good results, they ignore the context, which makes the inpainting images look less natural. Semantic image inpainting does better in this aspect. In this paper, we propose a new method about semantic image inpainting by improving the model of Generative Adversarial Networks. The generator of our network model refers to Boundary Equilibrium Generative Adversarial Networks and adds spectral normalization to improving unstable training. Then, our network chooses discriminator of Self-Attention Generative Adversarial Networks which is more concise. After training, the network model is limited to generate image encoding by context loss and prior loss to make it as similar as possible to the missing image encoding. Then, the generated images encoding corresponding to the without missing parts of the real images are obtained. After fusing the generated missing parts with the unmissed parts of the real image, it is the inpainting image. By the comparisons of vision and quantization, it shows that our method can well complete the task of large missing regions.