Structure-Prior-Guided GAN-Based Deep Learning for Blind Deconvolution and Optical Image Deblurring
Computer and Decision Making: An International Journal · 09.11.2026 · 3 (8S) · 978-1000
DOI: 10.1922/comdem.v3i8s.19237 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 09.11.2026. Verfasst von Sudharshan Banakar und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Blind image deblurring is a severely ill-posed task that remains a significant challenge because the blur kernel and the spatial variation of nonuniform motion are both unknown. Although existing deep learning-based methods have demonstrated impressive results in acquiring complex features using large datasets, they tend to be black boxes that do not consider the importance of classical image structure priors. Such disregard may result in an inefficient reconstruction of fine details and the appearance of additional unwanted artifacts. In this study, an SP–GAN, a structure prior-guided generative adversarial network, is proposed for blind nonuniform image deblurring. We utilize large amounts of image gradient data in a data-driven manner with our framework to reconcile deep feature learning with structural rigor. We apply a patch-wise Minimal Pixel prior to the GAN model to encourage generator reassignment of boundary sharpness and local sparsity conservation. In contrast to the conventional approaches that deconvolution and kernel estimation are independent processes, our multi-scale discriminator is designed and used to provide global and local consistency, with the use of Convolutional Neural Networks (CNNs). Experimental results show that the proposed SP–GAN outperforms other benchmark datasets, such as the GoPro and Kohler datasets, based on the peak signal-to-noise ratio and structural similarity index. These findings support the claim that the use of classical structure priors is effective in artifact suppression and improves the view as perceived in restored images over state-of-the-art CNN-based algorithms.
Zitieren
APA
Banakar, S., & K, C. (2026). Structure-Prior-Guided GAN-Based Deep Learning for Blind Deconvolution and Optical Image Deblurring. Computer and Decision Making: An International Journal, 3(8S), 978-1000. https://doi.org/10.1922/comdem.v3i8s.19237
Vancouver
Banakar S, K C. Structure-Prior-Guided GAN-Based Deep Learning for Blind Deconvolution and Optical Image Deblurring. Computer and Decision Making: An International Journal. 2026;3(8S):978-1000. doi:10.1922/comdem.v3i8s.19237
BibTeX
@article{Banakar2026,
author = {Banakar, Sudharshan and K, Chandrashekhar},
title = {Structure-Prior-Guided GAN-Based Deep Learning for Blind Deconvolution and Optical Image Deblurring},
journal = {Computer and Decision Making: An International Journal},
year = {2026},
volume = {3},
number = {8S},
pages = {978--1000},
doi = {10.1922/comdem.v3i8s.19237}
}
EUVIMED