结构-纹理分层下的并联去雾网络  

Multiple Dehazing Network Under Structure-Texture Decomposition

作　　者：任瑞琳 杨燕[1] REN Ruilin;YANG Yan(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730000,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730000

出　　处：《电光与控制》2024年第8期73-81,共9页Electronics Optics & Control

基　　金：国家自然科学基金(61561030);甘肃省优秀研究生“创新之星”项目(2023CXZX-547)。

摘　　要：针对去雾过程中的边缘丢失、颜色失真问题,提出一种基于结构-纹理分层的并联去雾网络。根据图像分层理论,将有雾图像分解为结构层和纹理层,搭建并联卷积神经网络;结构层图像含有大部分雾气和结构信息,可以使用深层网络进行去雾;纹理层图像含有纹理边缘等细节信息,可以使用浅层网络丰富纹理信息。网络采用多尺度卷积提升特征图的鲁棒性,跳跃连接减少运算的参数量。为避免图像重建中因误差产生的颜色失真,使用颜色可见度恢复模块进行颜色补偿。实验结果表明,所提算法对边缘细节的恢复有一定促进作用,能保留原有图像颜色,在客观指标和视觉效果上表现良好。Aiming at the problems of edge loss and color distortion in the process of dehazing,a multiple dehazing network based on structure-texture decomposition is proposed.According to the theory of image decomposition,the haze image is divided into structure layer and texture layer.Structure layer image contains most haze and structure information,a deep network is used for dehazing.The texture layer image contains detailed information such as texture edges,a shallow network is used to enrich the texture information.The network uses multi-scale convolution to improve the robustness of the feature map,and jump connection to reduce the number of parameters of the operation.To avoid color distortion caused by error in image reconstruction,the color visibility recovery module is used for color compensation.The experimental results show that the proposed algorithm promotes the restoration of edge details,retains the original image color,and performs well in objective indicators and visual effects.

关 键 词：图像去雾 图像分层 坐标注意力 纹理增强 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]