暗通道先验优化的生成对抗网络图像去雾算法  

Generative adversarial network image dehazing algorithm with a dark channel priori optimization

作　　者：苏腾华 吕莉[1,2] 樊棠怀[1,2] 谢海华 刘宝宏[1,2] SU Tenghua;LYU Li;FAN Tanghuai;XIE Haihua;LIU Baohong(School of Information Engineering,Nanchang Institute of Technology,Nanchang 330099,China;Nanchang Key Laboratory of IoT Sensing and Collaborative Computing for Smart City,Nanchang Institute of Technology,Nanchang 330099,China)

机构地区：[1]南昌工程学院信息工程学院,江西南昌330099 [2]南昌工程学院南昌市智慧城市物联感知与协同计算重点实验室,江西南昌330099

出　　处：《南昌工程学院学报》2025年第1期81-90,共10页Journal of Nanchang Institute of Technology

基　　金：国家自然科学基金资助项目(62463021)。

摘　　要：针对传统图像去雾方法存在的去雾图像失真、细节丢失且泛化性差等问题,提出一种暗通道先验优化的生成对抗网络图像去雾算法。首先,设计了一种新的模型框架,通过暗通道先验优化生成对抗网络,利用物理模型提高收敛性能;其次,采用残差自编码组成生成器网络,使用跳跃连接构成残差块保留图像细节信息;最后,引入马尔可夫判别器对去雾图像进行判别,反馈给生成器,进一步增强模型的去雾效果。在合成数据集以及真实数据集上进行训练测试,并在多种去雾场景下进行实验,结果表明该算法在多个数据集下的评价指标都为最高值,在多种去雾场景下均有良好表现,与传统暗通道先验算法相比性能提升了23%,并且该算法能够有效去除带雾图像中的雾层,较好还原图像细节内容,保证了较高的视觉质量。In order to solve the problems of dehazing image distortion,loss of detail and poor generalization of traditional image dehazing methods,this paper proposes a generative adversarial network image dehazing algorithm with a dark channel prior optimization.Firstly,a new model framework is designed,which generates an adversarial network through a priori optimization of dark channels,and the physical model is used to improve the convergence performance.Secondly,the residual auto-encoding is used to form a generator network,and the residual block is formed by jumping connections to retain the image detail information.Finally,the Markov discriminator was introduced to discriminate the dehazing image and fed back to the generator to further enhance the dehazing effect of the model.Experimental results show that the algorithm can effectively remove the fog layer in the foggy image,restore the image details well,ensure high visual quality,and perform well in a variety of dehazing scenarios.

关 键 词：图像去雾 生成对抗网络 暗通道先验 残差自编码 马尔可夫判别器 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]