Abstract—In recent years, deep learning technology is widely used in the field of facial image inpainting. The existing methods are prone to structural distortion, semantic capture blur and repair result distortion when dealing with large occluded areas. This paper proposes a facial image inpainting algorithm based on the Generative Adversarial Networks (GAN) framework, which combines the self-attention mechanism and the global local discriminator. First, the important features of the original image are captured by the attention layer in the generation network to obtain a larger receptive field and discard irrelevant information. In the discriminator network, the global and local discriminators are combined, and the good adaptability of the two discriminators to the overall semantics and local edge details is used to perform feedback training on the repair network and images. The final experimental results show that the algorithm in this paper has better repairing effect and higher training efficiency than other algorithms, and is superior to other existing algorithms in subjective visual perception and objective evaluation index.

Index Terms—facial image inpainting, generative adversarial networks, attention mechanisms, global and local discriminator, deep learning

Cite: Jianchu She and Ying Liu, "Facial Image Inpainting Algorithm Based on Attention Mechanism and Dual Discriminators," Journal of Image and Graphics, Vol. 10, No. 1, pp. 43-49, March 2022.

Copyright © 2022 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC BY-NC-ND 4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.