基于双向特征融合的输电线路异常目标检测  

Transmission line anomaly object detection based on two directions feature fusion

作　　者：田云龙 申贝贝 杜永杰 刘恒源 李辉[3] 陶冶[3] TIAN Yun-long;SHEN Bei-bei;DU Yong-jie;LIU Heng-yuan;LI Hui;TAO Ye(Haier Smart Home Digital Transformation Platform,Qingdao Haier Technology Co.,Ltd,Qingdao 266100,China;Strategic Development Center,National Engineering Research Center of Digital Home Networking,Qingdao 266100,China;School of Data Science,Qingdao University of Science and Technology,Qingdao 266061,China)

机构地区：[1]青岛海尔科技有限公司海尔智慧家数字化转型平台,山东青岛266100 [2]数字家庭网络国家工程研究中心战略发展中心,山东青岛266100 [3]青岛科技大学数据科学学院,山东青岛266061

出　　处：《计算机工程与设计》2024年第10期3051-3058,共8页Computer Engineering and Design

基　　金：山东省重点研发计划基金项目(重大科技创新工程)(2022ZDPT01)。

摘　　要：背景复杂、目标尺度变化大、数据集不均衡等是导致输电线路异常目标误检、漏检以及检测精度低的主要原因。因此,提出一种增强特征提取网络,有效减少特征提取过程中的信息丢失,更好保留小目标特征信息。使用通道优化与空间优化模块进行双向特征融合,以适应目标的多尺度变化,减少复杂背景信息的干扰。使用均衡采样与自适应类抑制损失,提高少数类别的检测精度,解决输电线路数据不平衡的问题。在输电线路异常目标检测任务中,检测精度达到90.5%,对困难场景有较好的检测效果。Complex backgrounds,large changes in target scales,many small targets,and unbalanced data sets are the main reasons for false detection,missed detection,and low detection accuracy of abnormal targets detection on transmission lines.Therefore,an enhanced feature extraction network was proposed,which effectively reduced the information loss during feature extraction and better retained small target feature information.Two directions feature fusion was performed using channel optimization and spatial optimization modules to adapt to the multi-scale changes of the target,and the interference of complex background information was reduced.The balanced sampling with adaptive class suppression loss was used to adaptively balance positive and negative sample gradients to improve the detection accuracy of a few classes and solve the problem of positive and negative samples as well as data imbalance in transmission line data sets.In the transmission line anomaly target detection task,the detection accuracy reaches 90.5%,which shows good detection effects for difficult scenarios.

关 键 词：输电线路 异常目标 目标检测 特征感知增强 双向特征融合 均衡采样 自适应类抑制损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]