基于图像分割和对抗训练的去雾算法  

Defogging Algorithm Based on Image Segmentation and Confrontation Training

作　　者：谢勇 贾惠珍 王同罕 雷初聪 徐铠珈 XIE Yong;JIA Huizhen;WANG Tonghan;LEI Chucong;XU Kaijia(Jiangxi Radioactive Geoscience Large Data Technology Engineering Laboratory,East China University of Technology,Nanchang 330000;Institute of Computer Science and Technology,Ningbo University,Ningbo 315211)

机构地区：[1]东华理工大学江西省放射性地学大数据技术工程实验室,南昌330000 [2]宁波大学计算机科学技术研究所,宁波315211

出　　处：《计算机与数字工程》2022年第8期1776-1781,1857,共7页Computer & Digital Engineering

基　　金：国家重点研发计划(编号:2018YFB1702700);江西省核地学数据科学与系统工程技术研究中心开放基金项目(编号:JETRCNGDSS201901);江西省教育厅科技项目(编号:GJJ190364)资助。

摘　　要：由于气候变化或者空气污染,不可避免地产生含雾图像。为了解决这一问题,论文提出了一种基于图像分割和对抗训练的去雾算法,在训练时,其首先将一张清晰图和一张雾图作为输入,随机进行线性或者非线性图像融合,得到一张含雾的混合图。其次,引入一个分离器将之再次分离成新的清晰图和雾图,利用两张原图和分离之后的两张分离图,利用交叉计算误差作为损失函数,对分离器进行训练优化,充分学习雾图特征。再者,引入一个鉴别器,对分离器处理完之后的图进行鉴别,鉴别混合图分离与否。在训练整个过程中,保证最小化分离器误差的同时,实现鉴别器鉴别率的最大化。在去雾时运用训练得到的模型,对雾图通过图像分割,从而达到去雾的目的。实验表明,在主观视觉效果和客观指标上,论文算法处于现存算法的前列,而且时间复杂度较低,实用性更强。Due to climate change or air pollution,foggy images are inevitably produced.In order to solve this problem,this pa⁃per proposes a defogging algorithm based on image segmentation and confrontation training.In training,firstly,it takes a clear im⁃age and a fog image as input,and randomly fuses linear or nonlinear images to get a mixed image containing fog.Secondly,a separa⁃tor is introduced to separate it into a new clear map and a fog map again.The separator is trained and optimized by using two original maps and two separated maps after separation,and the cross calculation error is used as the loss function to fully learn the character⁃istics of the fog map.Furthermore,a discriminator is introduced to discriminate the graph processed by the separator,and to dis⁃criminate whether the mixed graph is separated or not.In the whole training process,the separator error is minimized and the dis⁃crimination rate of the discriminator is maximized.When defogging,the trained model is used to segment the fog image,so as to achieve the purpose of defogging.Experiments show that the proposed algorithm is in the forefront of existing algorithms in terms of subjective visual effects and objective indicators,and has lower time complexity and stronger practicability.

关 键 词：去雾 图像分割 图像融合 分离器 鉴别器 

分 类 号：TN911[电子电信—通信与信息系统]