Occlusion Face Image Inpainting Method Based on Multi-scale and Contextual Attention
Chen Wang, Enzeng Dong, Sen Yang · 2021 China Automation Congress (CAC) · 2021
In order to solve the common issues of unclear repaired images and incoherent semantics in occluded face image inpainting, we propose a two-stage generation network that combines multi-scale joint and contextual attention. The first stage of the network uses a multi-scale parallel encoder-decoder to obtain preliminary inpainting results, and the second stage uses the context attention module to refine the inpainting details. Based on the idea of generating adversarial networks (GAN), we introduced a global&local discriminator to maintain the consistency of global and local content. Apart from the adversarial loss caused by GAN, our method also combines spatial attenuation reconstruction loss and a similar Markov random fields loss for training to further enhance the quality of the produced images. A large number of experiments on the CelebA and CelebA-HQ face datasets show that our method can obtain a relatively good inpainting effect.