AUTHOR=Lu Xiaoman , Lu Ran , Zhao Wenhao , Ma Erbin 

TITLE=Facial image inpainting for big data using an effective attention mechanism and a convolutional neural network

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 16 - 2022

YEAR=2023

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2022.1111621

DOI=10.3389/fnbot.2022.1111621

ISSN=1662-5218

ABSTRACT=Face image for big data is an important identity information of people. However, when it is transmitted in the public channel, there are often problems such as watermark occlusion, defilation and partial region missing. Therefore, it is of practical significance to study how to effectively restore the incomplete face image. In this paper, we propose a facial image inpainting method using multi-stage generative adversarial network and global attention mechanism. For the overall structure of the network, we use generative adversarial network as the main body then we establish skip connections to optimize network structure and use encoder-decoder structure to better capture semantic information of the missing part of the facial image. The local refinement network is proposed to enhance the local restore effect and weaken the influence of unsatisfactory results. Moreover, a global attention mechanism is added to the network to magnify the interactive features of the global dimension while reducing the information dispersion, which is more suitable for the restoration of human facial information. The comparative experiments on CelebA and CelebA-HQ big databsets show that the proposed method generate realistic inpainting results in both regular and irregular masks, and it can achive peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) and other evaluation indicators, which illustrate the performance and efficiency of the proposed model.