一种改进SSD的输电线路电力部件识别方法  

An improved SSD method for power component identification of transmission lines

作　　者：王朝硕[1] 李伟性[2] 郑武略 王宁[1] 赵航航 WANG Chaoshuo;LI Weixing;ZHENG Wulue;WANG Ning;ZHAO Hanghang(China Southern Power Grid EHV Transmission Company ICT Center,Guangzhou 510000,China;China Southern Power Grid EHV Transmission Company Guangzhou Bureau,Guangzhou 510000,China)

机构地区：[1]中国南方电网超高压输电公司信通中心,广东广州510000 [2]中国南方电网超高压输电公司广州局,广东广州510000

出　　处：《应用科技》2020年第4期75-81,共7页Applied Science and Technology

基　　金：中国南方电网有限责任公司科技项目(CGYKJXM20160006).

摘　　要：为解决单阶段多框检测器(single shot multibox setector,SSD)算法识别较小尺寸电力部件准确率低的问题,本文提出一种注意力机制和多尺度特征融合的单阶段多框检测器(attention mechanism and multiscale feature fusion single shot multibox detector,amSSD)算法。该方法在SSD网络特征提取层引入压缩和激励网络(squeeze-and-excitation networks,SENet)结构,筛选并保留更多与目标相关的特征通道;对浅层特征图采用膨胀卷积操作,使目标语义信息更加丰富;对高层特征图进行反卷积操作,并与浅层特征进行融合,得到具有更高分辨率高语义信息的目标特征图,提高对较小尺寸电力部件的识别能力。利用实际无人机飞行数据进行测试验证,实验结果表明:本文方法能够有效地识别出电力部件,而且识别平均准确率达到89.6%,比SSD方法的识别准确率提升了6.2%。In order to solve the problem of low accuracy of SSD(single shot multibox detector)algorithm in identifying small size power components,this paper proposes an attention mechanism and multi-scale feature fused amSSD(attention mechanism and multiscale feature fusion single shot multibox detector)algorithm.Firstly,the SENet(squeeze-and-excitation networks)structure is introduced into the feature extraction layer of SSD network to filter and retain more feature channels related to the target;secondly,the expansion convolution operation is adopted for the shallow feature map to enrich the semantic information of the target;finally,the deconvolution operation is carried out for the high-level feature map and the shallow feature is integrated.The target feature map with higher-resolution semantic information is obtained to improve the recognition ability of small size power components.The experiment has been carried out with actual UAV flight data.The results show that this method can effectively identify power components,and the average recognition accuracy is 89.6%,which is 6.2%higher than that of SSD method.

关 键 词：SSD模型 电力部件 多尺度 注意力机制 膨胀卷积 特征融合 反卷积 语义信息 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]