基于粗到细的多尺度单幅图像去雾方法  

Single-image dehazing via a coarse-to-fine multiscale approach

作　　者：王德文[1,2] 陈威 苏攀 WANG Dewen;CHEN Wei;SU Pan(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,China;Engineering Research Center of the Ministry of Education for Intelligent Computing of Complex Energy System,North China Electric Power University,Baoding 071003,China)

机构地区：[1]华北电力大学控制与计算机工程学院,河北保定071003 [2]华北电力大学复杂能源系统智能计算教育部工程研究中心,河北保定071003

出　　处：《智能系统学报》2024年第5期1102-1110,共9页CAAI Transactions on Intelligent Systems

基　　金：河北省自然科学基金项目(F2021502013);河北省高等学校科学技术研究项目(QN2023181);中央高校基本科研业务费专项资金项目(2021MS094).

摘　　要：为了解决现有图像去雾算法易出现细节纹理丢失、颜色失真或对非均匀浓雾处理不彻底的问题,提出一种基于粗到细的多尺度单幅图像去雾方法。首先,主干网络使用残差特征注意力模块对有雾图像进行特征提取;其次,将不同尺度的输入图像进行卷积预处理,通过多尺度特征融合模块将预处理的浅层特征与主干网络融合;再次,将不同粒度的非对称特征进行有效融合;最后,将浅层信息与深层信息自适应混合输出,通过对比正则损失构建正负样本信息,使得去雾图像更接近无雾图像。实验结果表明,与已有代表性的去雾方法相比,提出的方法能对合成数据集与真实数据集进行有效去雾,在细节保留与色彩还原上优于对比方法。A single-image dehazing method based on a coarse-to-fine,multiscale approach is proposed to address the shortcomings of existing dehazing methods,which often result in loss of detailed texture,color distortion,or incomplete processing of non-uniform dense fog.First,features are extracted from foggy images using the residual feature attention module of the backbone network.Second,images at different scales are preprocessed using convolution,and the preprocessed shallow features are fused with the backbone network through a multiscale feature fusion module.Asymmetric features at different levels are then effectively fused.Finally,shallow information is adaptively combined with deep information at the output.Compared with contrastive regularization loss,positive and negative sample information is constructed to make the defogged image more similar to a fog-free image.The experimental results show that,compared with existing representative defogging methods,the proposed method effectively defogs synthetic and real datasets and outperforms comparison methods in terms of detail retention and color reproduction.

关 键 词：图像去雾 粗到细 多尺度特征融合 残差特征注意力 非对称特征融合 自适应混合 对比正则 正负样本 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]