Face Image Super-Resolution Reconstruction Based on the Improved ESRGAN

Mingda Li

2023

Abstract

Researchers have investigated face super-resolution (SR) reconstruction extensively because it can effectively improve biometric identification, surveillance, and image improvement. This study's primary objective is to enhance generative adversarial networks (GANs) in order to create powerful face super-resolution models. Specifically, this paper introduces the improved enhanced SR-GAN (ESRGAN) model as a baseline. In addition, this paper proposes the Perceptual Image-Error Assessment through Pairwise Preference (PieAPP) loss function to fuse and perceive the visual quality of facial images. Secondly, eliminating the batch normalization layer is proposed to improve the model structure while incorporating residual dense blocks (RRDB) in the generator. The introduction of PieAPP loss can refine the SR process and emphasize the quality of the generated image. This study is conducted on an extensive and diverse face dataset, which includes facial images at various resolutions. Experimental results show that the "L1+Visual Geometry Group (VGG)+PieAPP" loss function is always better than the original ESRGAN model in various quantitative indicators. These improvements result in superior perceived image quality and user satisfaction. The practical significance of this research lies in its contribution to the development of more realistic and aesthetically pleasing facial reconstructions.

Paper Citation

in Harvard Style

Li M. (2023). Face Image Super-Resolution Reconstruction Based on the Improved ESRGAN. In Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-705-4, SciTePress, pages 471-476. DOI: 10.5220/0012820100003885

in Bibtex Style

@conference{daml23,
author={Mingda Li},
title={Face Image Super-Resolution Reconstruction Based on the Improved ESRGAN},
booktitle={Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2023},
pages={471-476},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012820100003885},
isbn={978-989-758-705-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML
TI - Face Image Super-Resolution Reconstruction Based on the Improved ESRGAN
SN - 978-989-758-705-4
AU - Li M.
PY - 2023
SP - 471
EP - 476
DO - 10.5220/0012820100003885
PB - SciTePress