一种改进条件生成对抗网络的图像去雾算法  

An Image Defogging Algorithm Based on an Improved Conditional Generation Adversarial Network

作　　者：薛子豪 王忠美 伍宣衡 李敏[1] 龙永红[1] XUE Zihao;WANG Zhongmei;WU Xuanheng;LI Min;LONG Yonghong(College of Railway Transportation,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学轨道交通学院,湖南株洲412007

出　　处：《湖南工业大学学报》2023年第4期50-55,共6页Journal of Hunan University of Technology

基　　金：湖南省自然科学基金资助项目(2020JJ5128);国家级创新创业课题基金资助项目(S202111535505,S202111535041);湖南工业大学教学改革课题基金资助项目(2020A14)。

摘　　要：针对现有去雾算法存在的颜色失真、细节丢失和一些肉眼可见的薄雾残留等问题,提出了一种改进条件生成对抗网络的去雾算法。首先,为了能更好地保留图像的底层纹理信息和结构信息,共享浅层和深层图像之间的特征,设计了含对称层跳跃连接结构的生成器。其次,为了保留图像的细节并减少伪影,重新设计了损失函数,在原始网络损失的基础上引入L1损失和感知损失。提出的算法在HSTS数据集上的峰值信噪比可达27.3064 dB,结构相似度达0.9633,比其他算法的最优值分别提高了5.728 dB和0.0581。去雾后目标检测的mAP提高了2.51%,召回率提高了4.31%。实验结果表明,所提出算法可以减少色差,解决薄雾残留问题,块效应基本消除,在主观效果和客观评价上均具有明显优势。In view of such flaws as color distortion,detail loss and visible haze residues found in existing defogging algorithms,a defogging algorithm has thus been proposed based on an improved conditional generation of adversarial network.First of all,for a better preservation of the underlying texture information and structure information of the image so as to share the features between the shallow and deep images,a generator with symmetric layer jump connection structure has thus been designed.Secondly,in order to preserve the details of the image and reduce the artifacts,the loss function is redesigned,with L1 loss as well as perceptual loss introduced based on the original network loss.The peak signal-to-noise ratio of the proposed algorithm on the HSTS data set can reach as high as 27.3064 dB,and the structural similarity can reach 0.9633,which is 5.728 dB and 0.0581 higher than the optimal values of other algorithms respectively.Target detection mAP after defogging has been improved 2.51%,meanwhile the recall rate has been improved 4.31%.The experimental results show that the proposed algorithm can help to reduce the color difference,remove the haze residue and basically eliminate the block effect,showing that it is characterized with obvious advantages in both subjective and objective evaluation.

关 键 词：条件生成对抗网络 图像去雾 感知损失 跳跃连接 

分 类 号：TN911.73[电子电信—通信与信息系统]