Abstract: Aiming at the problems of the local chromatic aberrations, boundary artifacts and detail defects in inpainting facial images with large damaged areas, this paper proposes a facial image inpainting model based on partial convolution and multi-scale feature integration. The model is divided into two components: a multi-scale inpainting network and a discriminator network. The inpainting network achieves feature extraction and integration of facial images by effectively fusing deep and shallow image features through a multi-level feature extraction module and a main branching module. Moreover, a joint loss function consisting of content loss, perceptual loss, style loss, total variance loss and adversarial loss is constructed for training a multi-scale inpainting network and enhancing visual consistency between generated images and real images through mutual adversarial with discriminator networks. Experimental results show that, under different mask ratio conditions, this model can generate images with reasonable texture structure and contextual semantic information, and perform better in qualitative and quantitative comparisons. 

Key words: facial image inpainting, partial convolution, multi-scale feature integration, generative adversarial network