A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction

The SRGAN model based on Generative Adversarial Networks (GANs) has achieved breakthroughs in perceptual quality for image super-resolution reconstruction. However, existing models suffer from feature smoothing effects in the generator and unstable training of the discriminator. To address these lim...

Full description

Saved in:
Bibliographic Details
Main Authors: Ximin Qu, Cheng Peng
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11075683/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850101699231875072
author Ximin Qu
Cheng Peng
author_facet Ximin Qu
Cheng Peng
author_sort Ximin Qu
collection DOAJ
description The SRGAN model based on Generative Adversarial Networks (GANs) has achieved breakthroughs in perceptual quality for image super-resolution reconstruction. However, existing models suffer from feature smoothing effects in the generator and unstable training of the discriminator. To address these limitations, we propose an improved SRGAN model (denoted as SRGAN-GBDB) through multi-strategy collaborative optimization to improve reconstruction quality. The specific improvements include: Firstly, in the generator, we remove batch normalization (BN) layers to alleviate feature degradation and embed the Convolutional Block Attention Module (CBAM) to enhance high-frequency feature modeling. Secondly, we implement spectral normalization for the discriminator and incorporate spectrally normalized residual blocks into the discriminator to strengthen its discriminative capability. Additionally, the joint perceptual-adversarial-content loss function of the SRGAN model balances reconstruction accuracy and visual quality. Experimental results show that the optimized model SRGAN-GBDB outperforms the traditional Bicubic interpolation method, SRGAN, and several other existing super-resolution algorithms on the Set5, Set14, BSD100, Urban100, and DIV2K datasets. It not only improves quantitative metrics such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) but also demonstrates superior visual quality through better detail recovery and fewer artifacts, validating the effectiveness and superiority of the optimization strategy. Future work will explore more efficient optimization methods and lightweight network architectures to facilitate the application of image super-resolution systems in mobile platforms and specialized fields like remote sensing and medical imaging.
format Article
id doaj-art-fa8ac6da61734c649f504c7355f3c3ce
institution DOAJ
issn 2169-3536
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-fa8ac6da61734c649f504c7355f3c3ce2025-08-20T02:39:56ZengIEEEIEEE Access2169-35362025-01-011311975711977110.1109/ACCESS.2025.358720111075683A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image ReconstructionXimin Qu0https://orcid.org/0009-0008-7998-2700Cheng Peng1https://orcid.org/0000-0003-0619-2771School of Computer Science and Technology, Xinjiang Normal University, Ürümqi, ChinaSchool of Computer Science and Technology, Xinjiang Normal University, Ürümqi, ChinaThe SRGAN model based on Generative Adversarial Networks (GANs) has achieved breakthroughs in perceptual quality for image super-resolution reconstruction. However, existing models suffer from feature smoothing effects in the generator and unstable training of the discriminator. To address these limitations, we propose an improved SRGAN model (denoted as SRGAN-GBDB) through multi-strategy collaborative optimization to improve reconstruction quality. The specific improvements include: Firstly, in the generator, we remove batch normalization (BN) layers to alleviate feature degradation and embed the Convolutional Block Attention Module (CBAM) to enhance high-frequency feature modeling. Secondly, we implement spectral normalization for the discriminator and incorporate spectrally normalized residual blocks into the discriminator to strengthen its discriminative capability. Additionally, the joint perceptual-adversarial-content loss function of the SRGAN model balances reconstruction accuracy and visual quality. Experimental results show that the optimized model SRGAN-GBDB outperforms the traditional Bicubic interpolation method, SRGAN, and several other existing super-resolution algorithms on the Set5, Set14, BSD100, Urban100, and DIV2K datasets. It not only improves quantitative metrics such as PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) but also demonstrates superior visual quality through better detail recovery and fewer artifacts, validating the effectiveness and superiority of the optimization strategy. Future work will explore more efficient optimization methods and lightweight network architectures to facilitate the application of image super-resolution systems in mobile platforms and specialized fields like remote sensing and medical imaging.https://ieeexplore.ieee.org/document/11075683/Optimization strategyimage super resolutionattention mechanismspectral normalizationresidual block
spellingShingle Ximin Qu
Cheng Peng
A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction
IEEE Access
Optimization strategy
image super resolution
attention mechanism
spectral normalization
residual block
title A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction
title_full A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction
title_fullStr A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction
title_full_unstemmed A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction
title_short A Study on an Improved Generative Adversarial Network Model for Super-Resolution Image Reconstruction
title_sort study on an improved generative adversarial network model for super resolution image reconstruction
topic Optimization strategy
image super resolution
attention mechanism
spectral normalization
residual block
url https://ieeexplore.ieee.org/document/11075683/
work_keys_str_mv AT ximinqu astudyonanimprovedgenerativeadversarialnetworkmodelforsuperresolutionimagereconstruction
AT chengpeng astudyonanimprovedgenerativeadversarialnetworkmodelforsuperresolutionimagereconstruction
AT ximinqu studyonanimprovedgenerativeadversarialnetworkmodelforsuperresolutionimagereconstruction
AT chengpeng studyonanimprovedgenerativeadversarialnetworkmodelforsuperresolutionimagereconstruction