The Analysis of Image Inpainting Based on Pix2Pix Model and Mix Loss

Xu Yan

2023

Abstract

This paper investigates the application of image translation as an important use case for generative adversarial networks (GAN), which has received widespread attention from scholars in recent years. This study builds an Image-to-image translation (pix2pix) model with a U-net as the generator and PatchGAN as the discriminator and observes the performance by adjusting the loss function of the generator. First experiment is modifying the scale factors for GAN Loss and Mean Absolute Error loss (L1 Loss), and the second is exchanging L1 Loss with square loss function (L2 Loss). After comparing the image authenticity and detail processing of different results, it is noticed that an overall better translation is achieved when the scale factor is set to 1:100. If finer detail handling is required, lowering the scale factor to 1:10 can be beneficial. However, it’s also found that including L2 Loss in the generator loss function do not yield favorable results. It provides guidance for future choices of hyperparameters for the pix2pix model and lays the foundation for further research into loss functions.

Paper Citation

in Harvard Style

Yan X. (2023). The Analysis of Image Inpainting Based on Pix2Pix Model and Mix Loss. In Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML; ISBN 978-989-758-705-4, SciTePress, pages 438-443. DOI: 10.5220/0012804900003885

in Bibtex Style

@conference{daml23,
author={Xu Yan},
title={The Analysis of Image Inpainting Based on Pix2Pix Model and Mix Loss},
booktitle={Proceedings of the 1st International Conference on Data Analysis and Machine Learning - Volume 1: DAML},
year={2023},
pages={438-443},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012804900003885},
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 - The Analysis of Image Inpainting Based on Pix2Pix Model and Mix Loss
SN - 978-989-758-705-4
AU - Yan X.
PY - 2023
SP - 438
EP - 443
DO - 10.5220/0012804900003885
PB - SciTePress