摘要
输电线路航拍图像存在背景复杂多变、检测目标占比较小的问题。针对部分图像属于阴影、模糊等视觉信息较差的困难样本,在特征融合角度的基础上,使用通道注意力使得模型更加关注复杂背景下的关键特征提取区域;基于自适应空间特征融合(Adaptively Spatial Feature Fusion,ASFF)机制使得浅层和深层的特征图更合理地融合;对检测模型的损失函数进行改进,解决损失函数无法准确反映真实框与预测框重合度大小的问题。在自建的金具目标检测数据集上进行实验,实验结果表明,所提出的改进算法在原始YOLOx-S(You Only Look Once x-S)基础上获得了5.15%的检测精度提升,召回率提高了1.62%,并且针对小目标、易漏检和错检目标的检测有了明显改善,体现了在输电线路上金具目标检测的优越性和实用性。
The aerial images of transmission lines have the problems of complex and changeable backgrounds with small proportion of detection targets.As some images are difficult samples with poor visual information such as shadows and blurring,based on feature fusion angle,channel attention is used to make the model pay more attention to the key feature extraction areas in complex backgrounds.Secondly,an adaptive feature fusion mechanism based on Adaptively Spatial Feature Fusion(ASFF)is used to make shallow and deep feature maps to fuse more reasonably.Finally,the loss function of the detection model is improved,in order to solve another problem that the loss function cannot accurately reflect the coincidence degree between the real box and the prediction box.The experiment is carried out on the self-built fitting target detection dataset.The experimental results show that the improved algorithm has achieved 5.15%improvement in detection accuracy based on the original YOLOx-S(You Only Look Once x-S),and the Recall value has increased by 1.62%.In addition,the detection of small targets,easily missed targets and mistakenly detected targets have been significantly improved,which reflecting the superiority and practicality of fitting target detection on transmission lines.
作者
赵振兵
吕雪纯
王帆帆
蒋志钢
张凌浩
杨迎春
ZHAO Zhenbing;LYU Xuechun;WANG Fanfan;JIANG Zhigang;ZHANG Linghao;YANG Yingchun(Department of Electronic and Communication Engineering,North China Electric Power University,Baoding 071003,China;Engineering Research Center of Intelligent Computing for Complex Energy Systems,Ministry of Education,North China Electric Power University,Baoding 071003,China;Hebei Key Laboratory of Power Internet of Things Technology,North China Electric Power University,Baoding 071003,China;Measuring Center of Sichuan Power Grid,Chengdu 610045,China;State Grid Sichuan Electric Power Research Institute,Chengdu 610095,China;Sate Grid Sichuan Electric Power Company,Chengdu 610041,China)
出处
《无线电工程》
北大核心
2023年第11期2664-2672,共9页
Radio Engineering
基金
国家自然科学基金(61871182,U21A20486)
河北省自然科学基金(F2020502009,F2021502008,F2021502013)。
关键词
输电线路巡检
金具检测
深度学习
特征融合
损失函数
注意力机制
transmission line inspection
fitting detection
deep learning
feature fusion
loss function
attention mechanism
