~摘要
鸟类会在一些输电杆塔上筑巢，形成鸟害，严重威胁输电线路的安全运行。为了解决鸟害问题，需要对输电线路进行巡检，检测鸟巢。目前的输电线路巡检已经进入了无人机巡检代替人工巡检的时代，并结合了机器视觉和深度学习技术，实现了自动巡线和检测。本文针对目前存在的输电线路巡检鸟巢检测算法速度较慢、精度较低、多目标和小目标检测效果较差的问题，提出了一种轻量级YOLOv3输电线路鸟巢检测方法，具体的主要内容有：
首先，针对由于天气原因导致的输电线路鸟巢巡检数据集图像质量较差的问题，通过双边平滑滤波进行图像去噪和弱化背景，并通过基于导向滤波的暗通道先验去雾方法进行图像去雾，来提升图像质量；针对数据量较少，无法构建数据集的问题，提出了基于C-DCGAN的数据增强方法和Mosaic数据增强方法来增强数据集。然后，针对当前鸟巢检测算法速度慢、检测精度不高的问题，本文对YOLOv3算法进行深入研究，设计了基于深度可分离卷积的主干网络，权重文件大小由241.082MB下降到45.374MB，FPS从原始模型的17.45提高到38.80；并设计了基于PANet的特征融合结构，经过官方数据集测试，精度大幅提升，mAP值从86.37%上升到了93.29%。在主干网络和特征融合结构优化的基础上，本文提出了基于深度可分离卷积和PANet的轻量化高精度鸟巢检测算法。最后，在实验测试阶段，采用了基于迁移学习和标签平滑的优化训练策略，在本文构建的输电线路鸟巢巡检数据集上进行模型训练。与原始模型的对比试验实现了本文算法在损失函数、mAP精度指标和F1精度指标方面的性能验证，并通过逐项消融实验和与其他算法的对比实验，验证本文算法改进的有效性和相比其他算法的优越性。
实验结果表明，本文提出的输电线路鸟巢检测算法能够对输电线路鸟巢进行有效检测，在检测精度上和检测速度上，优于同类算法，平均检测精度可达95%，在分辨率为416×416的图像中FPS达到36.91，可以胜任目前阶段的输电线路巡检图像鸟巢检测任务。

~Birds will nest on some transmission towers, forming bird damage, which seriously threatens the safe operation of transmission lines. In order to solve the problem of bird damage, it is necessary to inspect the transmission line and detect the bird nest. The current transmission line inspection has entered the era of unmanned aerial vehicle inspection instead of manual inspection, and combines machine vision and deep learning technology to achieve automatic line inspection and detection. Aiming at the problems of slow speed, low accuracy and poor detection effect of multi-target and small target in the current transmission line inspection nest detection algorithm, this paper proposes a lightweight YOLOv3 transmission line nest detection method. The main contents are as follows:
Firstly, aiming at the problem of poor image quality of transmission line bird nest inspection data set caused by weather, the image denoising and background weakening are carried out by bilateral smoothing filtering, and the image defogging is carried out by dark channel prior defogging method based on guided filtering to improve the image quality. Aiming at the problem of small amount of data and unable to build data sets, the data enhancement method based on C-DCGAN and Mosaic data enhancement method are proposed to enhance the data set.Then, in view of the slow speed and low detection accuracy of the current nest detection algorithm, this paper conducts in-depth research on the YOLOv3 algorithm, and designs the backbone network based on deep separable convolution. The weight file size decreases from 241.082MB to 45.374MB, and the FPS increases from 17.45 of the original model to 38.80. The feature fusion structure based on PANet was designed. After the official data set test, the accuracy was greatly improved, and the mAP value increased from 86.37% to 93.29%. Based on the optimization of backbone network and feature fusion structure, this paper proposes a lightweight and high-precision bird nest detection algorithm based on deep separable convolution and PANet. Finally, in the experimental test stage, the optimization training strategy based on transfer learning and label smoothing is adopted to train the model on the transmission line bird nest inspection data set constructed in this paper. Compared with the original model, the performance verification of the proposed algorithm in terms of loss function, mAP accuracy index and F1 accuracy index is realized, and the effectiveness of the proposed algorithm improvement and its superiority over other algorithms are verified by item-by-item ablation experiments and comparative experiments with other algorithms.
The experimental results show that the proposed transmission line nest detection algorithm can effectively detect the transmission line nest. In terms of detection accuracy and detection speed, it is superior to the similar algorithms. The average detection accuracy can reach 95%, and the FPS reaches 36.91 in the image with resolution of 416×416, which can be competent for the current stage of transmission line inspection image nest detection task.

[1]张涛,赵东艳,薛峰,等.电力系统智能终端信息安全防护技术研究框架[J].电力系统自动化,2019,43(19):1-8+67.

[2]陆建遵.电工技术基础[M].北京:中国电力出版社,2010:6.

[3]中国电力企业联合会.中国电力行业年度发展报告2020[R].北京:中国电力企业联合会,2020.

[4]李隆基,周文涛,李学刚,等.架空输电线路防鸟害技术措施[J].智慧电力,2016,44(04):95-98.

[5]王少华,叶自强.架空输电线路鸟害故障及其防治技术措施[J].高压电器,2011,47(2):61-67.

[6]许家文,阴酉龙,刘小双,等.输电线路人机协同巡检模式的研究与应用[J].电气时代,2021(05):24-29.

[7]邵瑰玮,刘壮,付晶,等.架空输电线路无人机巡检技术研究进展[J].高电压技术,2020,46(01):14-22.

[8]Fu Y, Zhang Z, Mi Y, et al. Droop Control for DC Multi-Microgrids Based on Local Adaptive Fuzzy Approach and Global Power Allocation Correction[J]. IEEE Transactions on Smart Grid, 2018: 1-1.

[9]Goodfellow, I., Bengio, Y., Courville, A. Deeplearning[J]. Cambridge: MITpress, 2016: 326-366.

[10]叶明武,钟超逸,张璐娟,等.基于人工神经网络的智能化架空线路应变快速解调方法[J].半导体光电,2022,43(01):188-194.

[11]包晓敏,王思琪.基于深度学习的目标检测算法综述[J].传感器与微系统,2022,41(04):5-9.

[12]Luis Mejías, Correa J F, Iván Fernando Mondragón, et al. COLIBRI: A vision-Guided UAV for Surveillance and Visual Inspection[C]// IEEE International Conference on Robotics & Automation. Rome, Italy: IEEE, 2007: 2760-2761.

[13]Montambault S, Beaudry J, Toussaint K, et al. On the application of VTOL UAVs to the inspection of power utility assets[C]// Applied Robotics for the Power Industry (CARPI), 2010 1st International Conference on. Montreal, Canada: IEEE, 2010: 1-7.

[14]黄仁超.基于无人机图像的输电线杆塔上的鸟巢检测技术[D].广东工业大学,2018.

[15]王桢.基于视频图像处理技术的输电线路杆塔鸟巢检测[D].华北电力大学,2013.

[16]徐晶,韩军,童志刚,等.一种无人机图像的铁塔上鸟巢检测方法[J].计算机工程与应用,2017,53(06):231-235.

[17]林庆达,禤亮,覃威威,等.电力系统图像检测研究综述[J].云南电力技术,2017,45(04):30-33.

[18]Nassim A, Haikel A, Yakoub B, et al. Deep Learning Approach for Car Detection in UAV Imagery[J]. Remote Sensing, 2017, 9(4): 312.

[19]Murugendrappa N, Gayathri K E, Harshitha G M, et al. Multiple Object detection using Autonomous UAVs based on YOLOv3 algorithm[J]. IOP Conference Series: Materials Science and Engineering, 2020, 925(1): 012025 (6pp).

[20]杨沛.双判别器生成对抗网络及其在接触网鸟巢检测的应用研究[D].西南交通大学,2018.

[21]王纪武,罗海保,鱼鹏飞,等.基于Faster R-CNN的多尺度高压塔鸟巢检测[J].北京交通大学学报,2019,43(05):37-43.

[22]NGUYEN K, HUYNH N T, NGUYEN P C, et al. Detecting Objects from Space: An Evaluation of Deep-Learning Modern Approaches[J]. Electronics, 2020, 9(4):583-583.

[23]时磊,杨恒,周振峰,等.基于RetinaNet模型的鸟巢智能检测[J].电力大数据,2020, 23(02):53-58.

[24]彭闯.输电线路无人机巡检图像中电力部件检测方法研究[D].重庆理工大学, 2020.

[25]杨波,曹雪虹,焦良葆,等.改进实时目标检测算法的电力巡检鸟巢检测[J].电气技术,2020,21(05):21-27+32.

[26]刘国文,张彩霞,李斌,等.基于改进RetinaNet模型的接触网鸟巢检测[J].数据采集与处理,2020,35(03):563-571.

[27]祁婕,焦良葆.改进SSD的输电铁塔鸟窝检测[J].计算机系统应用,2020,29(05):202-208.

[28]钟映春,孙思语,吕帅,等.铁塔航拍图像中鸟巢的YOLOv3检测研究[J].广东工业大学学报,2020,37(03):42-48.

[29]蒲天骄,乔骥,韩笑,等.人工智能技术在电力设备运维检修中的研究及应用[J].高电压技术,2020,46(02):369-383.

[30]陈凯.深度学习模型的高效训练算法研究[D].中国科学技术大学,2016.

[31]张军阳,王慧丽,郭阳,扈啸.深度学习相关研究综述[J].计算机应用研究,2018,35(07):1921-1928+1936.

[32]田娟,李英祥,李彤岩.激活函数在卷积神经网络中的对比研究[J].计算机系统应用,2018,27(07):43-49.

[33]周飞燕,金林鹏,董军.卷积神经网络研究综述[J].计算机学报,2017,40(06):1229-1251.

[34]GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation[C]// 2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus, USA: IEEE, 2014: 580-587.

[35]Girshick R. Fast RCNN[J]. Proceedings of the IEEE International Conference on Computer Vision, 2015: 1440-1448.

[36]Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Det-ection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 39(6): 1137-1149.

[37]Liu W, Anguelov D, Erhan D, et al. SSD: Single Shot MultiBox Detector[C]//European conference on computer vision. Springer, Cham, 2016: 21-37.

[38]Redmon J, Divvala S, Girshick R, et al. You Only Look Once: Unified, Real-Time Object Detection[C]// Computer Vision & Pattern Recognition. Las Vegas, USA: IEEE, 2016: 779-788.

[39]LIN T Y, P Dollár, GIRSHICK R, et al. Feature Pyramid Networks for Object Detection[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, USA: IEEE, 2016: 936-944.

[40]BAHMAIN B, MOSELEY B, VATTANI A, et al. Scalable K-Mean++[J]. Proceedings of the VLDB Endowment, 2012, 5(7): 622-633.

[41]NEUBECK A, Gool L. Efficient Non-Maximum Suppression[C]// International Conference on Pattern Recognition. Hong Kong, China: IEEE, 2006: 850-855.

[42]Redmon J, Farhadi A. YOLOv3: An IncrementalImprovement[EB/OL]. [2022-03-16]. http://arxiv.org/abs/1804.02767.

[43]李青君,孟庆昊,沈妍,等.图像平滑技术研究与应用[J].科技创新与应用,2022,12(02):165-167.

[44]靳阳阳,韩现伟,周书宁,等.图像增强算法综述[J].计算机系统应用,2021,30(06):18-27.

[45]He K M, Sun J, Tang X O. Single image haze removal using dark channel prior[C]// Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition. Miami, USA: IEEE, 2009: 1956-1963.

[46]Jean Pouget-Abadie, Mehdi Mirza, et. al. Generative Adversarial Netw-orks[J]. Machine Learning, 2014: 147-160.

[47]YUN S, HAN D, OH S J, et al. CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features[EB/OL]. [2022-03-17]. http://arxiv.org/abs/1905.04899.

[48]张姝,王昊天,董骁翀,等.基于深度学习的输电线路螺栓检测技术[J].电网技术,2021,45(07):2821-2829.

[49]HOWARD A G, ZHU M, CHEN B, et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications[EB/OL]. [2022-03-16]. http://arxiv.org/abs/1704.04861.

[50]LIU S, QI L, QIN H, et al. Path Aggregation Network for Instance Segmentation[C]// 2018 IEEE/CVF Conference on ComputerVision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018: 8759-8768.

[51]吴昊.基于合成数据集的图像处理深度学习方法研究[D].兰州大学,2021.

[52]Pan S J , Qiang Y. A Survey on Transfer Learning[J]. IEEE Transactions on Knowledge and Data Engineering, 2010, 22(10): 1345-1359.

[53]SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the Inception Architecture for Computer Vision[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, USA: IEEE, 2016: 2818-2826.

[54]Loshchilov I, Hutter F. SGDR: Stochastic Gradient Descent with Warm Restarts[C]// ICLR 2017 (5th International Conference on Learning Representations). Freiburg, Germany: ICLR, 2016.