基于数据增强和注意力机制的输电线异物检测算法的研究  

Research on Obstructive Object Detection Algorithm of Transmission Line Based on Data Enhancement and Attention Mechanism

作　　者：齐国营 高彦飞 李剑武 吴永华 QI Guoying;GAO Yanfei;LI Jianwu;WU Yonghua(New Energy Branch of Huaneng Shaanxi Power Generation Co.,Ltd.,Xi’an 710075,China)

机构地区：[1]华能陕西发电有限公司新能源分公司,陕西西安710075

出　　处：《微型电脑应用》2025年第2期55-60,共6页Microcomputer Applications

基　　金：国家自然科学基金项目(52061042,U1134106);中国华能集团有限公司课题项目(HNKJ21-HF252)。

摘　　要：输电线异物检测模型的训练数据集存在种类多和数据少的特点,而现有的方法在深度网络特征学习过程中存在细节特征丢失的情况,对此,提出了使用数据增强和增加注意力机制的单框多次检测器(SSD)框架对输电线异物进行检测。对采集的输电线异物图像进行预处理,主要包括对图像的颗粒噪声进行高斯去噪,然后进行直方图均衡化;使用Mosaic方式对输电线异物检测模型的训练数据集进行扩充,提高异物检测模型的鲁棒性和泛化能力;将注意力机制挤压—激励(SE)网络模块引入SSD检测框架,能够高效地学习不同Channel之间的特征,并进行特征融合,从而能够快速和精准提取关键的特征信息。试验结果表明,基于数据增强和注意力机制的输电线异物检测算法能够对输电线异物进行更加准确的检测,所提算法相较于Faster RCNN、SSD和YOLOv3检测算法提高了5个百分点、3个百分点和6个百分点,模型平均检测速度减小了0.021 s、0.007 s和0.003 s。There are many kinds but few data in the training data set of the transmission line obstructive object model,so that existing methods have the situation that the detailed features are lost in the process of deep network feature learning.In this paper,a single shot multibox detector(SSD)framework with data enhancement and attention enhancement mechanism is proposed to detect foreign bodies in power lines.The collected images of foreign bodies in power lines are preprocessed,which mainly includes Gaussian denoising of particle noise and then histogram equalization.The Mosaic method is used to expand the foreign body model training dataset of transmission lines to improve the robustness and generalization ability of the foreign body detection model.The attention mechanism squeeze-and-excitation(SE)network module is introduced into the SSD detection framework,which can efficiently learns the features between different channels and perform feature fusion,so as to extract the key feature information accurately and quickly.The testing results show that the proposed foreign body detection algorithm based on data enhancement and attention mechanism can detect foreign bodies in power lines more accurately.Compared with Faster RCNN,SSD and YOLOv3 detection algorithms,the detection speed of the model is improved by 5 percentage points,3 percentage points and 6 percentage points,and the detection speed of model is reduced by 0.021 s,0.007 s and 0.003 s,respectively.

关 键 词：输电线异物检测 Mosaic数据增强 SSD目标检测 注意力机制 SE网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]