A ResNet and U-Net Hybrid Discriminator Based on GANs for Facial Images Inpainting

Yue Zhu

2023

Abstract

Image inpainting technique is an application of computer vision technology, whose main task is to complete reconstructing image with missing or damaged areas by using the information of the existing part, which becomes a crucial part in computer vision (CV). This research presents a method for optimizing discriminators in Generative Adversarial Networks (GANs) inpainting application. By introducing a hybrid discriminator comprised by Residual Network-50 (Resnet50) and U-Net network, it has better perception of reconstructed facial image textures and regional contours, thereby reducing the common blurring problem of reconstructed images. Through random blurring images pre-training for the proposed discriminator, only the fine-tuned process is required in practical applications, resulting in faster training time. Through comparative experiments, this research can achieve smaller training losses, while the subjective image effect after reconstruction is also significantly improved, specifically in areas such as the nose and eyes, where the edge resource is more pronounced. The discriminator optimization design proposed in this research, combined with pre- training methods, can be applied to GANs image inpainting application, significantly improving the reconstruction effect of complex texture areas in damaged images, especially facial areas.

Paper Citation

in Harvard Style

Zhu Y. (2023). A ResNet and U-Net Hybrid Discriminator Based on GANs for Facial Images Inpainting. In Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-705-4, SciTePress, pages 192-196. DOI: 10.5220/0012819400003885

in Bibtex Style

@conference{daml23,
author={Yue Zhu},
title={A ResNet and U-Net Hybrid Discriminator Based on GANs for Facial Images Inpainting},
booktitle={Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2023},
pages={192-196},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012819400003885},
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 - A ResNet and U-Net Hybrid Discriminator Based on GANs for Facial Images Inpainting
SN - 978-989-758-705-4
AU - Zhu Y.
PY - 2023
SP - 192
EP - 196
DO - 10.5220/0012819400003885
PB - SciTePress