DWG-Network: Unsupervised Image Super-Resolution via Dynamic-Weighted Generation
Journal of Imaging · 20.11.2026 · 12 (11) · 590
DOI: 10.3390/jimaging12110590 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Imaging, online erschienen am 20.11.2026. Verfasst von Doudou Ma und 3 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Super-resolution (SR) maps low-resolution (LR) images to high-resolution (HR) images by enriching LR image details, and recent studies focus on learning the common detailed features between unpaired LR and SR images to supplement details for generating target HR images in an unsupervised learning manner. However, the learning process in existing studies tends to either over-guide the model towards unpaired LR images or over-prevent such over-guidance, resulting in low generation quality. In this paper, we propose a new method, named the “dynamic-weighted SR image generative network” (DWG-Network), to address this issue. Our DWG-Network consists of two encoder–decoders and three discriminators. First, we input both the LR images (ODLR) downsampled from the original HR and the unpaired LR (ULR) images into the first encoder to obtain the salient features and reconstruct these features back into LR images. Then, we feed the encoded features into the first discriminator to evaluate their similarity and find common features to complement the information lost during downsampling. Meanwhile, the second discriminator is employed to evaluate the detail similarity between the reconstructed ODLR and ULR images. Later, the reconstructed ODLR images are fed to another encoder–decoder to generate the target HR images. Lastly, we send both the generated and original HR images to the third discriminator to judge whether the current images are original or generated HR images. The details of the reconstructed ODLR image are learned from the ULR image, which can easily lead to over-guidance. The third discriminator will always judge the generated image as fake, which leads to over-prevention. Therefore, we set the weights of the two evaluation scores of the second and third discriminators as dynamic in the loss to balance over-guidance and over-prevention and ensure that the weighted values come from the previous round of the two discriminators. Experimental results on DIV2K, Set5, and Set14 demonstrate the competitive performance of DWG-Network against the evaluated unsupervised SR methods.
Zitieren
APA
Ma, D., Chen, G., Li, W., & Li, W. (2026). DWG-Network: Unsupervised Image Super-Resolution via Dynamic-Weighted Generation. Journal of Imaging, 12(11), 590. https://doi.org/10.3390/jimaging12110590
Vancouver
Ma D, Chen G, Li W, Li W. DWG-Network: Unsupervised Image Super-Resolution via Dynamic-Weighted Generation. Journal of Imaging. 2026;12(11):590. doi:10.3390/jimaging12110590
BibTeX
@article{Ma2026,
author = {Ma, Doudou and Chen, Guibin and Li, Wenhao and Li, Wei},
title = {DWG-Network: Unsupervised Image Super-Resolution via Dynamic-Weighted Generation},
journal = {Journal of Imaging},
year = {2026},
volume = {12},
number = {11},
pages = {590},
doi = {10.3390/jimaging12110590}
}
EUVIMED