基于改进U-Net结构生成对抗网络的运动模糊绝缘子图像复原方法  

Motion Blurred Insulator Image Restoration Method Based on Improved U-Net Structure Generative Adversarial Network

作　　者：崔昊杨 韩奕 张驯 王茺 刘诚 CUI Haoyang;HAN Yi;ZHANG Xun;WANG Chong;LIU Cheng(College of Electronics and Information Engineering,Shanghai University of Electric Power,Yangpu District,Shanghai 200090,China;Electric Power Research Institute of State Grid Gansu Electric Power Company,Lanzhou 730070,Gansu Province,China;School of Materials Science and Engineering,Yunnan University,Kunming 650091,Yunnan Province,China;Changsha Hongze Power Technology Co.,Ltd.,Changsha 410015,Hunan Province,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海市杨浦区200090 [2]国网甘肃省电力公司电力科学研究院,甘肃省兰州市730070 [3]云南大学材料科学与工程学院,云南省昆明市650091 [4]长沙市泓泽电力技术有限公司,湖南省长沙市410015

出　　处：《电网技术》2023年第6期2594-2603,共10页Power System Technology

基　　金：国家自然科学基金项目(52177185);国网甘肃省电力公司科技项目(52272220002U)。

摘　　要：为通过提高输电线路巡检航拍绝缘子过程中产生的运动模糊图像质量来提高检测准确率,文章提出一种基于改进U-Net结构生成对抗网络的运动模糊绝缘子图像复原方法。为有效融合图像深、浅层特征,基于U-NetL1结构搭建生成网络,并在特征融合过程添加并行注意力机制,加强对有效特征的学习,下采样层级联空洞空间金字塔池化与残差模块,扩宽感受野和加深网络深度的同时避免梯度消失问题;另外,采用内容损失与对抗损失作为损失函数,以提高复原图像质量。构建绝缘子运动模糊图像数据集进行实验,与Deblur GAN、SRN等方法的对比结果表明,本文方法的峰值信噪比、结构相似度指标均高于其他算法;采用YOLOv4进行目标检测实验,其结果表明该方法对提升绝缘子检测准确率具有一定实际意义。In order to improve the detection accuracy by increasing the motion blur image quality generated during the transmission of the transmission line inspection aerial insulation,a motion blur insulating sub-image restoration method based on improved U-Net structures generates a counterfeit network is proposed.In order to effectively fuse the deep and shallow features of the image,a generative network is built based on the U-NetL 1 structure,and a parallel attention module is added to the feature fusion process to strengthen the learning of effective features,while widening the receptive field and deepening the network depth,the gradient disappearance problem is avoided;in addition,content loss and adversarial loss are used as loss functions to improve the quality of restored images.The insulator motion blur data set is constructed for experiments,and the comparison results with DeblurGAN,SRN and other methods show that the peak signal-to-noise ratio and structural similarity indicators of this method are higher than other algorithms;YOLOv4 is used for target detection experiments,and the results show that this method can improve the accuracy of insulator detection.rate has a certain practical significance.

关 键 词：运动模糊图像复原 生成对抗网络 U-Net 注意力机制 空洞空间金字塔池化 

分 类 号：TM721[电气工程—电力系统及自动化]