摘要
现代化战争高度依赖图像等载体收集情报。雾天环境下得到的图像会干扰对战场场景的清晰呈现,还把重要特征隐匿其中,影响信息获取。针对目前图像去雾算法普遍存在颜色失真、图像细节丢失等问题,提出一种基于卷积神经网络(Convolutional Neural Network, CNN)与Transformer混合架构的多尺度特征交互去雾网络(Multi-scale Feature Interactive DehazeNet, MFI-DehazeNet),采用编解码结构,以端到端的方式实现单幅图像去雾。MFI-DehazeNet网络首先设计了一种多尺度特征交互模块,该模块实现了CNN网络特征的跨尺度融合;其次改进了Transformer结构,设计一种全局特征表达模块来增强整个网络的全局表达能力,用以弥补卷积结构感受野不足的问题;来自编码器的输出融合CNN和Transformer网络这两种异构架构的信息,该输出会通过特征重建模块(即解码器)进行处理,以恢复并重建出去雾后的图像。实验结果表明,相较于其他方法,MFI-DehazeNet无论是在合成有雾图像还是在真实有雾图像上都实现了更好的去雾效果。
Modern warfare highly relies on the carriers such as images to collect intelligence.The images obtained in foggy conditions can interfere with the clear presentation of a battlefield scene and also conceal the important features,thus affecting the acquisition of battlefield information.In view of the common issues,such as color distortion and image detail loss,of current image dehazing algorithms,this paper proposes a multi-scale feature interaction dehazing network(MFI-DehazeNet),which uses a hybrid architecture of convolutional neural network(CNN)and Transformers.The MFI-DehazeNet uses an encoder-decoder structure to achieve a single image dehazing in an end-to-end manner.First,a multi-scale feature interaction module that enables cross-scale fusion of CNN network features is introduced inMFI-DehazeNet.And then the Transformer structure is improved by using a global feature expression module to boost the network's global expression capability,thus addressing the receptive field limitations of convolutional structures.The output from the encoder,which integrates the two heterogeneous architectures of CNN and Transformer networks,is processed through the feature reconstruction module(i.e.,the decoder)to restore and reconstruct dehazed images.Experimental results indicate that MFI-DehazeNet outperforms other algorithms in dehazing both synthetic and real hazy images.
作者
刘昕昊
陈彬
应文健
李沛陶
伍世虔
LIU Xinhao;CHEN Bin;YING Wenjian;LI Peitao;WU Shiqian(School of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China;Naval University of Engineering,Wuhan 430033,Hubei,China)
出处
《兵工学报》
北大核心
2025年第5期201-212,共12页
Acta Armamentarii
基金
国家自然科学基金项目(62202347)。
作者简介
通信作者:应文健,邮箱:842812952@qq.com。
