Face Inpainting with Dilated Skip Architecture and Multi-Scale Adversarial Networks

Bin Jiang, Hongyu Liu, Chao Yang, Shangrong Huang, Yi Xiao

Abstract: Recent researches in neural network models have shown great potential in image inpainting task, which focus on finding reasonable area from known content to predict missing pixel values of incomplete image and generate new images. However they still create distorted structures and textures that are inconsistent with surrounding areas and require further post processing to optimize the results. Especially for face repair problems, it is often necessary to generate semantically new images that contain a large number of appearance. In addition, many recent approaches limited their applications in the field center region inpainting, which have difficulties to repair border regions or large missing area. Motivated by these observations, we introduce dilated skip architecture, which combines the advantages of dilated convolution and U-net. With this combination, our proposed network can enlarge the perception area for better repairing, as well as preserve the completeness of the original image features during the repairing process. Experiments on CelebA show that our approach can effectively deal with image inpainting problems and obtain images with higher qualities than previous ones.