摘要
提出了一种基于深度学习的图像去雾方法。利用一种深度估计和天空区域分割的方法合成并修正不同雾级的室外数据集;利用跳层连接和增加感知损失函数,提高雾图生成网络学习传播图的能力;利用判别网络对结果进行精修。结果表明,在合成雾图和自然航拍雾图上,提出的算法能较快地得到更加自然清晰的去雾图像,主客观评价优于其他算法。
An image dehazing method based on deep learning was proposed. A depth estimation and sky region segmentation method were used to synthesize and correct outdoor datasets with different fog levels. Using the skip layer connection and adding the perceptual loss function, the convolutional neurual network can generate transmission map more accurately. The discriminate network was used to refine the network’s parameters. The results show that on the synthetic and the natural aerial hazy image datasets, the proposed algorithm can quickly obtain superior performance than other representative algorithms in terms of objective indexs and the subjective visual quality.
作者
杨传栋
陈栋
刘桢
张晓龙
王昱翔
YANG Chuandong;CHEN Dong;LIU Zhen;ZHANG Xiaolong;WANG Yuxiang(Ammunition Technology Office, Army Academy of Artillary and Air Defense, Hefei 230031, China)
出处
《兵器装备工程学报》
CAS
北大核心
2019年第10期131-135,共5页
Journal of Ordnance Equipment Engineering
基金
军队“十三五”预研基金项目图像末制导××弹(301070103)
