基于幸运成像和生成对抗网络的大气湍流图像复原方法  

Atmospheric Turbulence Image Restoration Based on Lucky Imaging and Generative Adversarial Networks

作　　者：吕品 邓东平 石铁柱[1,2,3,4,5] 王梦迪 刘潜 田雨 张紫红 曾赟 邬国锋 Lyu Pin;Deng Dongping;Shi Tiezhu;Wang Mengdi;Liu Qian;Tian Yu;Zhang Zihong;Zeng Yun;Wu Guofeng(MNR Key Laboratory for Geo-Environmental Monitoring of Great Bay Area,Shenzhen University,Shenzhen 518060;Guangdong-Hong Kong-Macao Joint Laboratory for Smart Cities,Shenzhen 518060;State Key Laboratory of Subtropical Building and Urban Science,Shenzhen 518060;Guangdong Key Laboratory of Urban Informatics,Shenzhen University,Shenzhen 518060;Shenzhen Key Laboratory of Spatial Smart Sensing and Services,Shenzhen University,Shenzhen 518060)

机构地区：[1]深圳大学自然资源部大湾区地理环境监测重点实验室,深圳518060 [2]粤港澳智慧城市联合实验室,深圳518060 [3]亚热带建筑与城市科学全国重点实验室,深圳518060 [4]深圳大学广东省城市空间信息工程重点实验室,深圳518060 [5]深圳大学深圳市空间信息智能感知与服务重点实验室,深圳518060

出　　处：《计算机辅助设计与图形学学报》2025年第1期157-166,共10页Journal of Computer-Aided Design & Computer Graphics

基　　金：深圳市科技计划(ZDSYS20210623101800001);广东省科技创新战略专项资金(粤港澳联合实验室)(2020B1212030009).

摘　　要：在拍摄远距离目标时,视频序列图像受到大气湍流的影响从而产生畸变和模糊,为对视频序列大气湍流退化图像进行复原,提出了一种幸运成像与生成对抗网络相结合的方法.采用空域幸运成像方法,在有限的视频序列图像中挑选出幸运区域,将其拼接-排序后进行叠加,从而消除大气湍流带来的几何畸变;在此基础上引入DeblurGAN-v2模型,进一步提升图像质量.将高速相机拍摄的真实湍流退化图像作为研究对象,采用所提方法进行实验,并与图像重采样、灰度变换、巴特沃斯高通滤波、MPRNet模型和DeblurGAN模型等方法进行对比,并通过客观评价指标对不同方法的结果进行评估.实验结果表明,所提方法的Brenner梯度函数、Laplacian梯度函数、灰度差分函数(SMD)、熵函数(Entropy)、能量梯度函数(Energy)、PIQE以及Brisque指标相较于其他方法分别提升了194%,58%,84%,7%,55%,74%和163%.从主观效果上看,幸运成像与生成对抗网络相结合的方法能显著地提高图像的视觉质量,并有效地降低图像的模糊和几何畸变程度.When capturing distant targets,video sequence images are subject to atmospheric turbulence,resulting in distortion and blurring.To restore degraded images caused by atmospheric turbulence in video sequences,we propose an algorithm that combines lucky imaging with generative adversarial networks.The algorithm utilizes spatial lucky imaging to select fortunate regions from limited video sequence images,which are then stitched and sorted to eliminate geometric distortion induced by atmospheric turbulence.Additionally,the DeblurGAN-v2 model is introduced to further enhance image quality.Real turbulent degradation images captured by a high-speed camera are employed for experimentation using the proposed method.Comparative analyses are conducted with methods such as image resampling,grayscale transformation,Butterworth high-pass filtering,MPRNet model,and DeblurGAN model.Objective evaluation metrics are employed to assess the results of different algorithms.Experimental results indicate that the proposed method yields significant improvements in Brenner gradient function,Laplacian gradient function,SMD,entropy function,energy gradient function,PIQE,and Brisque indicators,showing enhancements of 194%,58%,84%,7%,55%,74%,and 163%,respectively,compared to other methods.From a subjective perspective,the algorithm combining lucky imaging with generative adversarial networks significantly enhances the visual quality of images and effectively reduces the degree of blurring and geometric distortion.

关 键 词：图像复原 大气湍流效应 幸运成像 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]