注意力残差密集网络的单幅图像去雾算法  

Single Image Defogging Algorithm Based on Attention Residual Dense Networks

作　　者：黄小芬[1] 林丽群[2] 卢宇 HUANG Xiaofen;LIN Liqun;LU Yu(Department of Information Technology,Concord University College Fujian Normal University,Fuzhou 350117,China;College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China)

机构地区：[1]福建师范大学协和学院信息技术系,福建福州350117 [2]福州大学物理与信息工程学院,福建福州350108

出　　处：《福建师范大学学报（自然科学版）》2023年第1期68-74,共7页Journal of Fujian Normal University：Natural Science Edition

基　　金：福建省高校产学合作项目(2021H6026);福建省中青年教师教育科研项目(JAT210649);福建师范大学协和学院智能计算与应用团队项目(2020-TD-001)。

摘　　要：针对现有去雾方法色彩失真、去雾不彻底、细节丢失等问题,提出一种模块化的端到端的单幅图像深度去雾网络.首先,利用多尺度卷积核对输入有雾图像提取充分的关键特征;其次,构建由残差密集块及上、下采样单元形成的行和列的网格网络结构,行列之间通过一种新颖的注意力机制进行特征融合与提取;最后,由残差密集块和卷积层构成的后处理模块进一步减少去雾图像的残余伪影.定量和定性实验结果表明,所提方法去雾性能优越.In order to solve the problems of color distortion, incomplete defogging and detail loss in the existing defogging methods, this paper proposes a modular end-to-end single-image defogging network. Firstly, sufficient key features are extracted from the foggy images using multi-scale convolutional kernels;secondly, a grid network structure formed by residual dense blocks and transition-up and transition-down cells is constructed, and feature fusion and extraction are performed between the ranks of the grid network by a novel attention mechanism;finally, a post-processing module composed of residual dense blocks and convolutional layers further reduces the residual artifacts of the defogged images. Quantitative and qualitative experimental results show that the proposed method has superior defogging performance.

关 键 词：图像去雾 多尺度 网格网络 注意力机制 密集连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]