引用本文:张萌,郭一娜,王海东,上官宏.基于YUV颜色空间和图卷积残差网络的图像去模糊算法[J].控制理论与应用,2026,43(3):691~696.[点击复制]
ZHANG Meng,GUO Yi-na,WANG Hai-dong,SHANGGUAN-Hong.Image deblurring algorithm based on YUV color space and graph convolutional residual network[J].Control Theory & Applications,2026,43(3):691~696.[点击复制]
基于YUV颜色空间和图卷积残差网络的图像去模糊算法
Image deblurring algorithm based on YUV color space and graph convolutional residual network
摘要点击 532  全文点击 68  投稿时间：2023-08-09  修订日期：2025-10-10
查看全文  查看/发表评论  下载PDF阅读器   HTML
DOI编号  10.7641/CTA.2024.30537
  2026,43(3):691-696
中文关键词  图像去模糊  空间结构  YUV颜色空间  图卷积网络(GCN)  图结构  深度学习
英文关键词  imagedeblurring  spatial structure  YUV color space  graph convolutional network (GCN)  graph structure  deep learning
基金项目  国家自然科学青年基金项目(62271341),山西省科技创新人才团队项目(202204051001018),山西省回国留学人员科研教研资助项目(HGKY201 9080, 2020–127), 山西省研究生优秀创新项目(2021Y679,2022Y689), 太原科技大学研究生教育创新项目(BY2023010)资助.
作者单位E-mail
张萌* 太原科技大学电子信息工程学院 b202215110036@stu.tyust.edu.cn 
郭一娜 太原科技大学电子信息工程学院  
王海东 太原科技大学电子信息工程学院  
上官宏 太原科技大学电子信息工程学院  
中文摘要
      图像去模糊需要在保留空间细节的同时确保高层次的上下文信息的平衡.针对模糊图像中的空间结构破 坏, 上下文信息扭曲以及RGB图像中的通道间强相关性造成的颜色不平衡等问题,本文提出一种基于YUV颜色空 间和图卷积网络(GCN)的图像去模糊算法(YUVGCR).首先,设计了用于图像去模糊的YUV与RGB颜色空间转换算 法, 以解决RGB通道间强相关性的问题.然后,利用GCN可以将特征图转换为预生成图的顶点,对特征图进行图卷 积, 从而合成构建图结构的数据.通过这样做,可以隐式地将图拉普拉斯正则化应用于特征图,使其更加结构化.实 验表明,YUVGCR的峰值信噪比(PSNR)为36.21dB,比先进算法提高了2.93dB.可视化去模糊结果可以看出,YUV GCR能产生更清晰的边缘和细节,图像去模糊的整体性能获得较大提升.
英文摘要
      Image deblurring requires a balance between preserving spatial details and maintaining high-level contex tual information. To address spatial structure degradation, contextual distortion, and color imbalance caused by strong inter-channel correlation in blurry images, this paper proposes a novel image deblurring algorithm called YUV graph con volutional residual network (YUVGCR), based on the YUV color space and graph convolutional network (GCN). Firstly, a YUV-RGB color space transformation algorithm is designed to mitigate the issue of robust inter-channel correlations in RGBchannels. Subsequently, utilizing GCN, feature maps can be mapped to vertices of a pre-generated graph, and graph convolutions can be applied to these feature maps, thereby synthesizing and constructing graph-structured data. Through this process, implicit graph Laplacian regularization can be employed on the feature maps, enhancing their structural orga nization. Experimental results demonstrate that YUVGCR achieves a PSNR of 36.21 dB, which is a 2.93 dB improvement over state-of-the-art algorithms. Visualizations of the deblurred results show that YUVGCR produces sharper edges and f iner details, significantly enhancing the overall performance of image deblurring.