基于改进YOLO算法的输电线小目标金具轻量化视觉识别方法  

Light Weight Visual Recognition Method of Small Target Hardware for Transmission Lines Based on Improved YOLO Algorithm

作　　者：邹德华 张宏伟[3] 江维 龚闯 ZOU Dehua;ZHANG Hongwei;JIANG Wei;GONG Chuang(Hunan Province Key Laboratory of Intelligent Live Working Technology and Equipment(ROBOT)(State Grid Hunan Extra-High Voltage Transmission Company),Changsha 420100,China;Live Inspection and Intelligent Operation Technology State Grid Corporation Laboratory(State Grid Hunan Extra-High Voltage Transmission Company),Changsha 420100,China;School of Mechanical Engineering and Automation,Wuhan Textile University,Wuhan 430073,China)

机构地区：[1]智能带电作业技术及装备(机器人)湖南省重点实验室(国网湖南省电力有限公司超高压输电公司),湖南长沙420100 [2]带电巡检与智能作业技术国家电网公司实验室(国网湖南省电力有限公司超高压输电公司),湖南长沙420100 [3]武汉纺织大学机械工程与自动化学院,湖北武汉430073

出　　处：《湖南电力》2025年第1期136-143,共8页Hunan Electric Power

基　　金：湖北省自然科学基金项目(2018CFB273);国网湖南省电力有限公司科技项目(SGHNSJ000ZJS2400739);湖北省教育厅科研计划项目(B2019067)。

摘　　要：针对基于深度学习的输电线金具检测方法在提高小目标检测精度的同时无法保证模型轻量化的问题,提出一种输电线小目标金具轻量化检测方法。该方法在YOLOv4的基础上,首先根据小目标数据特点,结合基于密度的聚类算法(density-based spatial clustering of applications with noise,DBSCAN)和多次K均值的锚框优化策略优化预设框选择,加快网络收敛,提高检测精度。其次,使用GhostNetV2实现模型轻量化。然后设计小目标加强特征提取网络(small object enhanced multi-scale detection network,SMD-Net)提高目标特征提取能力,加强浅层特征与深层特征融合,解决下采样跨步卷积导致的特征丢失问题。最后,使用焦点损失函数优化样本分配,并使用深度可分离卷积(depth-wise separable convolution,DSC)降低模型复杂度。在自建小金具数据集上该算法的平均精度均值(mAP)达64.73%,模型计算量和参数量比改进前分别降低了78%和71%,而检测精度几乎没有损失。在公开数据集VisDrone2019上mAP达38.43%,与其他算法相比,该算法具有更优的小目标检测性能。Aiming at the problem that deep learning-based transmission line hardware detection methods can not ensure the lightweight model while improving the detection accuracy of small targets,a lightweight detection method for small target hardware on transmission lines is proposed.Based on YOLOv4,the method firstly combines density-based spatial clustering of applications with noise and the anchor frame optimization strategy with multiple K-means to optimize the preset frame selection according to the characteristics of the small target data,speed up the network convergence and improve the detection accuracy.Secondly,GhostnetV2 is used to achieve model light weighting.Then the network of small object enhanced multi-scale detection network,is designed to enhance the feature extraction ability,strengthen the fusion of shallow and deep features,and solve the problem of feature loss caused by down sampling cross-step convolution.Finally,focal loss function is used to optimize sample allocation and depth-wise separable convolution is used to reduce model complexity.The mean average precision(mAP)of this algorithm on the self-constructed small hardwaredataset reaches 64.73%,and the model computation and parameter counts are reduced by 78%and 71%compared to the pre-improvement period,while there is almost no loss of detection accuracy.The mAP reaches 38.43%on the public dataset VisDrone2019,which gives the algorithm a superiorsmall target detection performance compared to other algorithms.

关 键 词：输电线金具 密度聚类 小目标检测 轻量化 特征融合 焦点损失函数 

分 类 号：TM726[电气工程—电力系统及自动化]