With the emergence of Generative Adversarial Networks, great progress has been made in image inpainting. However, most existing methods can produce plausible results, but fail to generate finer textures and structures. This is mainly due to the fact that (1) the generation of finer content in the masked region of an image is not constrained enough during network training, and (2) many different alternative pixels are exist to fill in the masked regions, making it very difficult for the inpainting network to generate reasonable sharp edges. To address these issues, we propose a Scale Enhanced GAN (SE-GAN) model which combines the constraints of large- and small-scale receptive fields of our tailor-made discriminators to achieve fine-grained constraint on image details, a novel edge loss to further ensure the sharpness of the generated image. Experiments on multiple datasets including faces(CelebA-HQ), textures(DTD), buildings(Facade) and natural images(ImageNet, Places2) show that our approach can generate higher quality inpainting results with more details than previous methods.

随着生成对抗网络的出现，图像修复已经取得了很大的进步。但是，大多数现有方法可以产生合理的结果，但无法生成更精细的纹理和结构。这主要是由于以下事实：（1）在网络训练期间，图像的被遮罩区域中较细的内容的生成没有受到足够的约束；（2）存在许多不同的替代像素来填充被遮罩区域，从而使其修复网络很难生成合理的锐利边缘。为了解决这些问题，我们提出了一种比例增强GAN（SE-GAN）模型，该模型结合了我们量身定制的鉴别器的大范围和小范围接收场的约束条件，以实现对图像细节的细粒度约束，一种新颖的边缘损失以进一步确保所生成图像的清晰度。