The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. ex. Some numerals are expressed as "XNUMX".
Copyrights notice
The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. Copyrights notice
Cet article traite de la super-résolution (SR) à image unique basée sur les réseaux de neurones convolutifs (CNN). On sait que la récupération des composantes haute fréquence dans les images SR de sortie des CNN apprises par les moindres erreurs carrées ou les moindres erreurs absolues est insuffisante. Pour générer des composants haute fréquence réalistes, les méthodes SR utilisent des réseaux contradictoires génératifs (GAN), composés d'un générateur et un expert discriminateur, Sont développés. Cependant, lorsque le générateur tente d'induire une erreur de jugement du discriminateur, non seulement les composantes haute fréquence réalistes mais aussi certaines artefacts sont générés et les indices objectifs tels que le PSNR diminuent. Pour réduire les artefacts dans les méthodes SR basées sur GAN, nous considérons l'ensemble de toutes les images SR dont les erreurs quadratiques entre les résultats de la réduction d'échelle et l'image d'entrée se situent dans une certaine plage, et proposons d'appliquer la projection métrique sur ceci ensemble cohérent dans les couches de sortie des générateurs. La technique proposée garantit la cohérence entre les images SR de sortie et les images d'entrée, et les générateurs avec la projection proposée peuvent générer des composantes haute fréquence avec peu d'artefacts tout en conservant celles à basse fréquence en fonction du niveau de bruit connu. Des expériences numériques montrent que la technique proposée réduit les artefacts inclus dans les images SR originales d'une méthode SR basée sur GAN tout en générant des composants haute fréquence réalistes avec de meilleures valeurs PSNR dans les deux cas. sans bruit et à la bruyant situations. Puisque la technique proposée peut être intégrée dans différents générateurs si le processus de réduction d'échelle est connu, nous pouvons donner de la cohérence aux méthodes existantes avec les images d'entrée sans dégrader les autres performances SR.
Hiroya YAMAMOTO
Ritsumeikan University
Daichi KITAHARA
Ritsumeikan University
Hiroki KURODA
Ritsumeikan University
Akira HIRABAYASHI
Ritsumeikan University
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copier
Hiroya YAMAMOTO, Daichi KITAHARA, Hiroki KURODA, Akira HIRABAYASHI, "Image Super-Resolution via Generative Adversarial Networks Using Metric Projections onto Consistent Sets for Low-Resolution Inputs" in IEICE TRANSACTIONS on Fundamentals,
vol. E105-A, no. 4, pp. 704-718, April 2022, doi: 10.1587/transfun.2021EAP1038.
Abstract: This paper addresses single image super-resolution (SR) based on convolutional neural networks (CNNs). It is known that recovery of high-frequency components in output SR images of CNNs learned by the least square errors or least absolute errors is insufficient. To generate realistic high-frequency components, SR methods using generative adversarial networks (GANs), composed of one generator and one discriminator, are developed. However, when the generator tries to induce the discriminator's misjudgment, not only realistic high-frequency components but also some artifacts are generated, and objective indices such as PSNR decrease. To reduce the artifacts in the GAN-based SR methods, we consider the set of all SR images whose square errors between downscaling results and the input image are within a certain range, and propose to apply the metric projection onto this consistent set in the output layers of the generators. The proposed technique guarantees the consistency between output SR images and input images, and the generators with the proposed projection can generate high-frequency components with few artifacts while keeping low-frequency ones as appropriate for the known noise level. Numerical experiments show that the proposed technique reduces artifacts included in the original SR images of a GAN-based SR method while generating realistic high-frequency components with better PSNR values in both noise-free and noisy situations. Since the proposed technique can be integrated into various generators if the downscaling process is known, we can give the consistency to existing methods with the input images without degrading other SR performance.
URL: https://global.ieice.org/en_transactions/fundamentals/10.1587/transfun.2021EAP1038/_p
Copier
@ARTICLE{e105-a_4_704,
author={Hiroya YAMAMOTO, Daichi KITAHARA, Hiroki KURODA, Akira HIRABAYASHI, },
journal={IEICE TRANSACTIONS on Fundamentals},
title={Image Super-Resolution via Generative Adversarial Networks Using Metric Projections onto Consistent Sets for Low-Resolution Inputs},
year={2022},
volume={E105-A},
number={4},
pages={704-718},
abstract={This paper addresses single image super-resolution (SR) based on convolutional neural networks (CNNs). It is known that recovery of high-frequency components in output SR images of CNNs learned by the least square errors or least absolute errors is insufficient. To generate realistic high-frequency components, SR methods using generative adversarial networks (GANs), composed of one generator and one discriminator, are developed. However, when the generator tries to induce the discriminator's misjudgment, not only realistic high-frequency components but also some artifacts are generated, and objective indices such as PSNR decrease. To reduce the artifacts in the GAN-based SR methods, we consider the set of all SR images whose square errors between downscaling results and the input image are within a certain range, and propose to apply the metric projection onto this consistent set in the output layers of the generators. The proposed technique guarantees the consistency between output SR images and input images, and the generators with the proposed projection can generate high-frequency components with few artifacts while keeping low-frequency ones as appropriate for the known noise level. Numerical experiments show that the proposed technique reduces artifacts included in the original SR images of a GAN-based SR method while generating realistic high-frequency components with better PSNR values in both noise-free and noisy situations. Since the proposed technique can be integrated into various generators if the downscaling process is known, we can give the consistency to existing methods with the input images without degrading other SR performance.},
keywords={},
doi={10.1587/transfun.2021EAP1038},
ISSN={1745-1337},
month={April},}
Copier
TY - JOUR
TI - Image Super-Resolution via Generative Adversarial Networks Using Metric Projections onto Consistent Sets for Low-Resolution Inputs
T2 - IEICE TRANSACTIONS on Fundamentals
SP - 704
EP - 718
AU - Hiroya YAMAMOTO
AU - Daichi KITAHARA
AU - Hiroki KURODA
AU - Akira HIRABAYASHI
PY - 2022
DO - 10.1587/transfun.2021EAP1038
JO - IEICE TRANSACTIONS on Fundamentals
SN - 1745-1337
VL - E105-A
IS - 4
JA - IEICE TRANSACTIONS on Fundamentals
Y1 - April 2022
AB - This paper addresses single image super-resolution (SR) based on convolutional neural networks (CNNs). It is known that recovery of high-frequency components in output SR images of CNNs learned by the least square errors or least absolute errors is insufficient. To generate realistic high-frequency components, SR methods using generative adversarial networks (GANs), composed of one generator and one discriminator, are developed. However, when the generator tries to induce the discriminator's misjudgment, not only realistic high-frequency components but also some artifacts are generated, and objective indices such as PSNR decrease. To reduce the artifacts in the GAN-based SR methods, we consider the set of all SR images whose square errors between downscaling results and the input image are within a certain range, and propose to apply the metric projection onto this consistent set in the output layers of the generators. The proposed technique guarantees the consistency between output SR images and input images, and the generators with the proposed projection can generate high-frequency components with few artifacts while keeping low-frequency ones as appropriate for the known noise level. Numerical experiments show that the proposed technique reduces artifacts included in the original SR images of a GAN-based SR method while generating realistic high-frequency components with better PSNR values in both noise-free and noisy situations. Since the proposed technique can be integrated into various generators if the downscaling process is known, we can give the consistency to existing methods with the input images without degrading other SR performance.
ER -