基于特征增强的双重注意力去雾网络  

Dual-Attention Dehazing Network Based on Feature Enhancement

作　　者：陈海秀[1,2] 黄仔洁 陆康 陆成 何珊珊[1] 房威志 卢海涛 陈子昂 CHEN Haixiu;HUANG Zijie;LU Kang;LU Cheng;HE Shanshan;FANG Weizhi;LU Haitao;CHEN Ziang(Nanjing University of Information Science&Technology,School of Automation,Nanjing 210000,China;Nanjing University of Information Science&Technology,Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology(CICAEET),Nanjing 210000,China)

机构地区：[1]南京信息工程大学自动化学院,南京210000 [2]南京信息工程大学、江苏省大气环境与装备技术协同创新中心,南京210000

出　　处：《电光与控制》2025年第1期15-20,67,共7页Electronics Optics & Control

基　　金：国家自然科学基金(61302189)。

摘　　要：针对现有去雾方法处理的图像细节模糊和色彩偏差等问题,提出了一种基于特征增强的双重注意力去雾网络。该网络采用编码器-解码器结构,设计了一个双重注意力特征增强模块,其中,利用Ghost模块替代非线性卷积,实现模型轻量化处理,通过RFB充分融合不同尺度的特征,实现均匀去雾,引入双重注意力实现信息跨通道与空间交互,保证模型性能和抑制噪声特征。使用RESIDE数据集对网络进行训练和测试。实验结果表明,所提算法在主观视觉和客观评价指标上均有优异表现,能有效地提升网络的特征提取能力,实现对不同场景雾图的色彩恢复,增强图像的对比度和清晰度。In order to solve the problems of detail blurring and color deviation of images processed by the existing dehazing methods,a dual-attention dehazing network based on feature enhancement is proposed.In this network,an encoder-decoder structure is used to design a dual-attention feature enhancement module,in which the Ghost module is used to replace the nonlinear convolution to realize the lightweight processing of the model.The Receptive Field Block(RFB)fully integrates the characteristics of different scales.Dual attention mechanism is introduced to realize cross-channel and spatial interaction of information,so as to ensure the performance of the model and suppress the noise features.The RESIDE dataset is used for network training and testing.The experimental results show that the proposed algorithm has excellent performance in both subjective visual and objective evaluation indicators,which can effectively improve the feature extraction ability of the network,realize the color restoration of foggy images in different scenes,and enhance the contrast and clarity of the image.

关 键 词：图像去雾 特征增强 并行分支结构 多尺度映射 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]