一种基于深度学习的两阶段图像去雾网络  

A TWO-STAGE IMAGE DEHAZING NETWORK BASED ON DEEP LEARNING

作　　者：吴嘉炜 余兆钗 李佐勇[1,2] 刘维娜 张祖昌[1,2] Wu Jiawei;Yu Zhaochai;Li Zuoyong;Liu Weina;Zhang Zuchang(College of Computer and Control Engineering,Minjiang University,Fuzhou 350121,Fujian,China;Fujian Provincial Key Laboratory of Information Processing and Intelligent Control(Minjiang University),Fuzhou 350121,Fujian,China;Department of Computer Engineering,Fujian Polytechnic of Information Technology,Fuzhou 350003,Fujian,China)

机构地区：[1]闽江学院计算机与控制工程学院,福建福州350121 [2]福建省信息处理与智能控制重点实验室(闽江学院),福建福州350121 [3]福建信息职业技术学院计算机工程系,福建福州350003

出　　处：《计算机应用与软件》2020年第4期197-202,共6页Computer Applications and Software

基　　金：国家自然科学基金面上项目(61772254);福建省科技厅引导性项目(2017H0030);福建省高校青年自然基金重点项目(JZ160467);福建省中青年教师教育科研项目(JT180406);闽江学院校长基金项目(103952018240)。

摘　　要：雾是户外图像降质的主要因素之一,图像去雾旨在恢复有雾图像中的内容。基于图像底层特征和先验知识的传统算法去雾效果不稳定。针对以上问题,受深度学习理论的启发,提出一种端到端的两阶段去雾深度神经网络算法。将图像去雾和图像超分辨率重建相结合,先利用编码器-解码器预测低分辨率雾霾残留图像,再利用亚像素卷积和残差块重建出原始分辨率雾霾残留图像,最后预测出原始分辨率无雾图像。在合成和真实有雾图像上的实验表明,该算法在定量评价和定性评价中均优于对比算法。Fog is one of the main factors of outdoor image degradation,and image dehazing is designed to recover content in foggy images.The traditional algorithm based on image underlying features and prior knowledge has an unstable dehazing effect.In view of the above problems,inspired by the deep learning theory,this paper proposes an end-to-end two-stage dehazing depth neural network algorithm.It combined image dehazing and image super-resolution reconstruction.The encoder-decoder was used to predict low-resolution haze residual image,and then the original haze residual image was reconstructed with sub-pixel convolution and residual block.Finally,we predicted the original resolution haze-free image.Experiments on synthetic and real foggy images show that our algorithm outperforms the comparison algorithm in quantitative evaluation and qualitative evaluation.

关 键 词：图像去雾 深度学习 超分辨率重建 编码器解码器 残差块 图像恢复 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]