基于暗通道先验引导的图像去雾网络  

Image dehazing network based on dark channel prior guidance

作　　者：黄淑英 夏钰锟 杨勇[3] 万伟国 邱根莹 HUANG Shuying;XIA Yukun;YANG Yong;WAN Weiguo;QIU Genying(School of Software,Tiangong University,Tianjin 300387,China;School of Information Management,Jiangxi University of Finance and Economics,Nanchang 330032,China;School of Computer Science and Technology,Tiangong University,Tianjin 300387,China;School of Software and Internet of Things Engineering,Jiangxi University of Finance and Economics,Nanchang 330032,China)

机构地区：[1]天津工业大学软件学院,天津300387 [2]江西财经大学信息管理学院,南昌330032 [3]天津工业大学计算机科学与技术学院,天津300387 [4]江西财经大学软件与物联网工程学院,南昌330032

出　　处：《北京航空航天大学学报》2024年第9期2717-2726,共10页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(61862030,62072218,62261025);江西省自然科学基金(20192ACB20002,20192ACBL21008);江西省博士后科研项目(2020KY44)。

摘　　要：基于深度学习的去雾方法多数直接学习有雾图像和无雾图像之间的映射关系,未结合有雾图像自身特点,存在雾信息检测不精确、去雾不彻底的问题。针对该问题,提出一种基于暗通道先验引导的图像去雾网络(DCPDNet)。通过卷积层提取有雾图像的浅层特征;构建2个特征增强模块(FEB)来增强图像的空间特征,该模块在2个尺度上对图像特征进行增强,即利用深层特征图实现语义特征的增强,浅层特征图对实现图像细节特征的增强;为使提取的特征更关注雾的区域,基于有雾图像中雾的成像特点设计基于引导图的特征校正模块(FCB),利用暗通道先验理论构建引导图将网络学习的注意力引导到有雾区域,对提取的深层特征图做进一步的细化和校正;利用残差结构的跳转连接,将增强的浅层特征补充网络丢失的细节特征,并经过卷积操作重建去雾后图像。实验结果证明:DCPDNet可以在保持模型轻量型及运行速度较快的情况下实现良好的去雾效果。与近年先进的去雾方法进行比较,DCPDNet不仅在效率上占有优势,其去雾效果在主观视觉感受和客观评价结果上都获得了更好的效果。Currently,the majority of dehazing techniques that utilize deep learning focus on directly acquiring the mapping relationship between a foggy image and a non-fog image.Because of the lack of combination with the characteristics of fog images,there are problems such as inaccurate detection of fog information and incomplete dehazing.In order to address the aforementioned issues,this study introduces a novel approach called the dark channel prior-guided image dehazing network(DCPDNet),which operates in an end-to-end manner.First,the shallow features of the input foggy image are extracted by several convolution layers.Secondly,two feature enhancement blocks(FEB)are constructed to enhance the spatial features of the image,which can enhance the image features on two scales,that is,the deep feature map is used to enhance the semantic features,and the shallow feature map is used to enhance the image details.Thirdly,in order to make the extracted features pay more attention to the fog area,a feature correction block(FCB)based on the guidance map is designed by considering the imaging characteristics of fog in the fog image.The FCB uses the dark channel theory to build a guidance map to guide the attention of network learning to the fog area,and further refine and correct the extracted deep feature map.Finally,by using the skip connection of the residual structure,the enhanced shallow features are used to supplement the details lost in the network,and the image after dehazing is reconstructed through several convolution operations.Multiple trials have demonstrated that DCPDNet is capable of achieving a satisfactory dehazing effect while maintaining a lightweight model and quick execution.Compared with some advanced dehazing methods proposed in recent years,DCPDNet has better performance in terms of efficiency,subjective visual perception and objective evaluation results.

关 键 词：图像去雾 暗通道先验 引导图 残差学习 特征增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]