一种基于FFA-Net改进的单幅图像去雾算法  

An improved single-image dehazing algorithm based on FFA-Net

作　　者：何钦 徐望明[1,2] 王义焕 罗扬 王薇 HE Qin;XU Wangming;WANG Yihuan;LUO Yang;WANG Wei(College of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;Engineering Research Center for Metallurgical Automation and Detecting Technology of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China;College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China)

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430081 [2]武汉科技大学冶金自动化与检测技术教育部工程研究中心,湖北武汉430081 [3]武汉科技大学计算机科学与技术学院,湖北武汉430065

出　　处：《武汉科技大学学报》2024年第6期448-456,共9页Journal of Wuhan University of Science and Technology

基　　金：国家自然科学基金项目(62202347);湖北省中央引导地方科技发展专项(2023EGA001);武汉科技大学冶金自动化与检测技术教育部工程研究中心开放课题(MADTOF2021B02).

摘　　要：雾霾会严重影响使用卷积神经网络的视觉系统对目标图像的检测和识别能力,为此本文在特征融合注意力网络FFA-Net的基础上设计和添加全局空间上下文增强(GSCE)模块和细节渐进增强(PDE)模块,进而提出一种改进型单幅图像去雾算法。GSCE模块用于增强全局空间信息,PDE模块用于逐步细化和增强图像特征,二者结合进行高效和轻量级的特征提取,弥补原模型中大量使用跳跃连接所造成的细节信息损失。改进模型分别在公共基准数据集RESIDE的室内数据和室外数据上进行训练,并分别在SOTS的室内和室外两个数据集上进行了测试。结果表明,本文算法明显超越了原FFA-Net和现有典型的单幅图像去雾算法,尤其在SOTS室内测试数据集上,单独融合GSCE模块就使得PSNR指标从36.36 dB提升到38.39 dB,在进一步使用PDE模块后PSNR指标提升到38.78 dB,算法的去雾性能得到较大提高,验证了改进策略的有效性。Haze can severely affect the ability of visual systems using convolutional neural networks to detect and recognize target images.To address this issue,an improved single-image dehazing algorithm was proposed by adding a global spatial context enhancement(GSCE)module and a progressive detail enhancement(PDE)module based on the feature fusion attention network(FFA-Net).The GSCE module was designed to enhance global spatial information,and the PDE module to progressively refine and enhance the image features.The combination of these two modules enabled efficient and lightweight feature extraction,compensating for the loss of detail information caused by the extensive use of skip connections in the original model.The improved model was trained on both indoor and outdoor data of the public benchmark dataset RESIDE and tested on the indoor and outdoor datasets of SOTS.Experimental results indicate that the proposed algorithm significantly outperforms the original FFA-Net and several existing typical single-image dehazing algorithms,especially on the SOTS indoor test dataset,where the PSNR metric increases from 36.36 dB to 38.39 dB by simply integrating GSCE module,and further increases to 38.78 dB with the use of PDE module.This demonstrates the great improvement in dehazing performance and validates the effectiveness of the designed strategies.

关 键 词：图像去雾 卷积神经网络 FFA-Net 全局空间上下文增强模块 细节渐进增强模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]