Image inpainting for irregular holes based on generative adversarial networks

Image inpainting aims to recover the missing regions in an image and reconstruct a satisfactory restoration result with high quality. To solve the problem that the existing image inpainting methods do not deal with the information of missing and non-missing regions flexibly, and the global and local restoration semantics are inconsistent, we design a three-stage restoration model for the different semantic information required for restored regions at different scales, which utilizes different sizes of receptive fields to provide better image details at multiple scales, including global and local, and ensures semantic consistency of contextual information. Experimenting with our method on three popular publicly available image drawing datasets, the results show that this paper's method outperforms current restoration models.