Efficient GAN-based Face Completion Method Design

Despite significant progress in deep learning-based face-related tasks, severe occlusion remains a challenging problem. In this paper, we investigate the use of Generative Adversarial Networks (GAN) to achieve high-quality face completion with a small training expenditure. We design an efficient GAN structure, introduce channel attention for the decoder, and design a multi-scale discriminator. Experiments show that under the same training conditions, our method exhibits superior results to similar methods: not only does it have higher quality of completion, but the model also converges faster.