HSVConnect: HSV guided enhanced content generation network for image inpainting

With the introduction of structural information, image inpainting technology has made great progress in inpainting large missing areas. However, most existing methods still suffer from structural distortion and blurred details when inpainting large corrupted areas. This is mainly due to the fact that (1) existing structural images struggle to completely and accurately represent global structure of image, and (2) with the deepening of encoding network, shallow texture features are lost step by step, which makes encoder struggle to use shallow texture features to restore fine texture. To address these issues, A novel two-stage inpainting network HSVConnect (HSVC) is proposed. In the first stage, HSVC uses HSV images to train a structure generator to reconstruct structure of missing area. In the second stage, a content generator containing a multi-scale encoding network is designed to provide different levels of texture features for decoder generation fine-grained details. Moreover, a content discriminator containing a multi-scale attention discrimination module is designed to enhance discriminant ability of content discriminator. Experiments on CelebA-HQ, Paris StreetView, and Places2 datasets show the superior performance of HSVC in inpainting large areas of damage. The HSVC test code and models are available at https://github.com/IPCSRG/HSVC-Inpainting .