Super-Resolution Using Deep Learning Focusing on Local Reproducibility

Detailed information, such as textures, in images can be challenging to reproduce owing to their relatively small size. Therefore, we propose a super-resolution method that focuses on small regions to reproduce texture. The method consists of three components: spatial similarity loss, a discriminato...

Full description

Saved in:
Bibliographic Details
Published in2023 IEEE 12th Global Conference on Consumer Electronics (GCCE) pp. 731 - 732
Main Authors Ishikawa, Mizuki, Seo, Masataka
Format Conference Proceeding
LanguageEnglish
Published IEEE 10.10.2023
Subjects
Correlation
Deep learning
Focusing
Image segmentation
Multitasking
Reproducibility of results
spatial similarity
super-resolution
Superresolution
Online AccessGet full text

Cover

More Information
Summary:Detailed information, such as textures, in images can be challenging to reproduce owing to their relatively small size. Therefore, we propose a super-resolution method that focuses on small regions to reproduce texture. The method consists of three components: spatial similarity loss, a discriminator, and multitask learning of super-resolution and segmentation. Spatial similarity loss reproduces the texture by bringing the correlation, not the individual pixel values, closer to the target. The discriminator assesses whether the generated image effectively reproduces the desired texture. The accuracy of texture reproduction was improved by applying these two methods to each small region, rather than applying them to the entire image.
ISSN:2693-0854
DOI:10.1109/GCCE59613.2023.10315493