基于多尺度编码-解码神经网络的图像去雾算法  

Image dehazing algorithm based on multiscale encoding decoding neural network

作　　者：王勇[1] 边宇霄 李新潮 徐椿明 彭刚 王继奎 WANG Yong;BIAN Yu-xiao;LI Xin-chao;XU Chun-ming;PENG Gang;WANG Ji-kui(College of Communication Engineering,Jilin University,Changchun 130012,China;Changchun Shikai Science and Technology Industry Co.,Ltd.,Changchnu 130015,China)

机构地区：[1]吉林大学通信工程学院,长春130012 [2]长春师凯科技产业有限责任公司,长春130015

出　　处：《吉林大学学报(工学版)》2024年第12期3626-3636,共11页Journal of Jilin University:Engineering and Technology Edition

基　　金：吉林省科技厅重点研发项目(20230201043GX)。

摘　　要：针对有雾的场景中图像采集系统采集到的图像会出现细节缺失、色彩暗淡、亮度降低的问题,在一体化去雾网络(AOD)理论的基础上提出一种多尺度编码-解码神经网络(MSAOD)模型进行图像去雾。本文网络模型分为3个模块:(1)预处理模块,将输入图像分为2部分进行预处理;(2)主干模块,通过多尺度编码器-解码器对第1部分的输出进行特征提取;(3)后处理模块,对特征图进行映射操作。通过训练得到去雾图像,实验结果表明,本文方法要优于主流的深度学习和传统方法的图像去雾效果,去雾后的图像在细节、色彩和亮度等方面都有所优化。Aiming at the problems of missing details,dim color and reduced brightness in the images collected by the image acquisition system in the foggy scene,a multi-scale encoding decoding neural network(MSAOD)model was proposed based on the AOD theory for image dehazing.The proposed network model was divided into three modules.The first module is the preprocessing module,which divides the input image into two parts for preprocessing.The second module is the backbone module,which extracts the features of the output of the first part through the multi-scale encoder decoder.The third module is the post-processing module,which maps the feature map.The experimental results show that this method is superior to the mainstream deep learning and traditional methods in image dehazing effect,and the image after dehazing is optimized in detail,color,brightness and so on.

关 键 词：数字图像处理 图像去雾 大气散射模型 深度学习 多尺度网络 

分 类 号：TP399[自动化与计算机技术—计算机应用技术] TN919.8[自动化与计算机技术—计算机科学与技术]