摘要
针对传统的基于先验知识与假设条件的去雾算法在实际应用中受到太多限制的问题,本文提出了一种基于端到端卷积神经网络的去雾算法,即通过学习雾天图像与清晰图像之间的映射关系实现图像去雾。首先,该算法采用了多尺度映射,通过多尺度的卷积提取出雾霾图像的较多细节信息特征;其次,运用反卷积以减少训练网络的复杂性;最后,结合浅层与深层的合并机制将删除特征图中的伪像素,提高恢复无雾图像的质量。实验结果表明,本文提出的去雾算法在自然雾天图像和合成雾天图像上均优于其它对比算法,并且合成雾天图像在结构相似度(SSIM)和峰值信噪比(PSNR)两个重要的图像评价指标上都取得了良好的表现。
Aiming at the problem that the traditional fogging algorithm,which is mostly based on prior knowledge and hypothesis can hardly be satisfied in practice,a kind of end-to-end convolutional neural network is proposed,by learning the mapping relationship between foggy images and clear images,image defogging can be realized directly.First,the algorithm is adopted by multi-scale mapping.Through multiscale convolution,it can extract haze features with more detailed information.Secondly,deconvolution is used to reduce the complexity of computer training network.Finally,Combining the the parallel algorithm of shallow layer and deep layer,the pseudo-pixels in the feature map will be deleted so as to improve the quality of image restoration without fog.The experimental results show that the proposed algorithm is superior to other algorithms in both natural fog images and composite fog images,and the composite fog images have achieved good performance on two important image evaluation indexes of SSIM and PSNR.
作者
陈清江
张雪
柴昱洲
CHEN Qing-jiang;ZHANG xue;CHAI yu-zhou(School of Science,Xi′an University of Architecture and Technology,Xi′an 710055,China;Xi′an Institute of Space Radio Technology,Xi′an 710000,China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2019年第2期220-227,共8页
Chinese Journal of Liquid Crystals and Displays
基金
国家自然科学基金(No.61403298)
陕西省自然科学基金(No.2015JM1024)
陕西省教育厅专项科研计划(No.2013JK0586)
陕西省教育厅自然科学基金(No.15JK2157)~~
关键词
卷积神经网络
多尺度映射
反卷积
大气散射模型
激活函数
convolution neural network
multi-scale mapping
deconvolution
atmospheric scattering Model
activation function
