[1]黄亚群,罗 俊*,蒋慕蓉,等.结合 GAN 和风格迁移的太阳斑点图重建方法[J].计算机技术与发展,2023,33(05):49-55.[doi:10. 3969 / j. issn. 1673-629X. 2023. 05. 008]
　HUANG Ya-qun,LUO Jun*,JIANG Mu-rong,et al.Reconstruction Method of Solar Speckle Image Combined with GAN and Style Transfer[J].,2023,33(05):49-55.[doi:10. 3969 / j. issn. 1673-629X. 2023. 05. 008]

点击复制

结合 GAN 和风格迁移的太阳斑点图重建方法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年05期

页码:

49-55

栏目:

媒体计算

出版日期:

2023-05-10

文章信息/Info

Title:

Reconstruction Method of Solar Speckle Image Combined with GAN and Style Transfer

文章编号:

1673-629X(2023)05-0049-07

作者:

黄亚群¹ ; 罗 俊^1* ; 蒋慕蓉¹ ; 杨 磊² ; 郑培煜¹

1. 云南大学 信息学院,云南 昆明 650500;

2. 中国科学院云南天文台,云南 昆明 650011

Author(s):

HUANG Ya-qun1 ; LUO Jun1* ; JIANG Mu-rong1 ; YANG Lei2 ; ZHENG Pei-yu1

1. School of Information Science and Engineering,Yunnan University,Kunming 650500,China;

2. Yunnan Observatories,Chinese Academy of Sciences,Kunming 650011,China

关键词:

太阳斑点图; 超分辨率重建; 生成对抗网络; 风格迁移; 深度学习

Keywords:

solar speckle image; super-resolution reconstruction; generative adversarial networks; style transfer; deep learning

分类号:

TP391. 41

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 05. 008

摘要:

在云南天文台拍摄的模糊太阳斑点图的超分辨率重建过程中,采用现有深度学习算法存在高频信息难以恢复、重建图不够清晰等问题,为此,提出一种结合 GAN(生成对抗网络) 和风格迁移网络的

太阳斑点图超分辨率重建方法 STYLE-NICE-GAN。 首先,利用 GAN 获取低分辨率太阳斑点图到 Level1+高分辨率太阳斑点图的映射关系,重建太阳斑点图的全局轮廓和部分细节;其次,使用风格迁移网络,对 GAN 的重建结果进行二次重建,在保留局部细节、高频信息和不影响后续分析的同时,提高图像的整体对比度和清晰度。 实验结果表明,与现有深度学习超分辨率重建算法相比,该方法具有重建图像清晰度更高、高频信息恢复能力更强的优点,重建结果在两个有参考评价指标 PSNR、SSIM 和三个无参考评价指标BRISQUE、NIQE、PIQE 上的评价均占有优势。

Abstract:

In the process of super-resolution reconstruction of the blur solar speckle images taken by the Yunnan Observatory,the existingdeep learning algorithms have problems such as difficulty in recovering high-frequency information and lack of clarity in reconstructedimages. Therefore,a super - resolution reconstruction method of solar speckle image named STYLE - NICE - GAN?

is proposed, whichcombines GAN ( Generative Adversarial Networks) and style transfer networks. Firstly,GAN is used to obtain the mapping relationshipfrom the low-resolution solar speckle image to the high-resolution solar speckle image and reconstruct the global outline and some detailsof the solar speckle image. Secondly,style transfer networks is used to perform secondary reconstruction on the reconstructed results ofGAN,which can improve the overall contrast and clarity of the image while preserving local details and high frequency information andnot affecting subsequent analysis. The experimental results show that compared with the existing deep learning super - resolutionreconstruction algorithms,the proposed method has the advantages of higher image definition and stronger ability of high-frequency information recovery,its reconstruction results are superior in the evaluation of two reference evaluation indicators PSNR,SSIM and three non-reference evaluation indicators BRISQUE,NIQE and PIQE.

常用功能