基于多尺度生成对抗网络的大气湍流图像复原  

Atmospheric Turbulence Image Restoration Based on Multi-Scale Generative Adversarial Network

作　　者：甄诚 杨永胜[1] 李元祥[1] 钟娟娟 ZHEN Cheng;YANG Yongsheng;LI Yuanxiang;ZHONG Juanjuan(School of Aeronautics and Astronautics,Shanghai Jiao Tong University,Shanghai 200240,China;AVIC Leihua Electronic Technology Research Institute,Wuxi,Jiangsu 214063,China)

机构地区：[1]上海交通大学航空航天学院,上海200240 [2]中国航空工业集团公司雷华电子技术研究所,江苏无锡214063

出　　处：《计算机工程》2021年第11期227-233,共7页Computer Engineering

基　　金：工业和信息化部民机专项(MJZ-2016-S-44)。

摘　　要：大气湍流会导致图像发生畸变和模糊,为对单幅大气湍流退化图像进行复原,提出一种基于多尺度生成对抗网络(GAN)的图像复原方法。采用GAN框架,在生成器网络中添加多尺度注意力特征提取单元与多层次特征动态融合单元,从而提升模型的感受野范围。在此基础上,引入特征融合机制以实现湍流退化图像复原。实验结果表明,相比标准GAN、SIU-Net模型,多尺度GAN能显著提高图像的视觉质量,有效降低图像的模糊和几何畸变程度。Atmospheric turbulence can produce distortions and blurs in images.To restore a single image affected by atmospheric turbulence,an image restoration method is proposed based on a multi-scale Generative Adversarial Network(GAN).Based on the GAN architecture,the method introduces units for the extraction of multi-scale attention features and dynamic fusion of multi-level features to enlarge the receptive field.On this basis,a feature fusion mechanism is added to implement restoration of images affected by atmospheric turbulence.The experimental results show that compared with the standard GAN and the SIU-Net model,the proposed multi-scale GAN can significantly improve the visual quality of images,and reduce the blurs and geometric distortions.

关 键 词：图像复原 图像重建 生成对抗网络 大气湍流效应 多尺度特征 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]