轻量级网络算法对输电线路上异物目标的识别

doi:10.16180/j.cnki.issn1007-7820.2023.04.010

摘要/Abstract

摘要：

针对输电线路上多种异物所引起的电力巡检问题,可以采用深度学习图像识别方法进行检测。文中提出了一种改进型轻量级网络检测算法模型,通过将YOLOv4的主干特征提取网络替换为轻量级神经网络GhostNet,减少图片输入时计算所产生的特征图冗余;对YOLOv4的PANet模块进行修改,采用深度可分离卷积模块替换其中的普通卷积模块,可减轻参数计算量。结果表明,相比于原YOLOv4检测算法,该改进型算法在IOU阈值取0.5时,平均精准度下降2.1%,但检测速度达到了原算法的2.21倍,参数计算量仅为原算法的17.84%。与其他几种算法的对比表明新算法的参数指标表现满足需求。在维持较高精确度的情况下,文中所提算法的检测速度得到提升,计算量减少,证明了其在目标检测时的有效性与可行性。

关键词: 输电线路, 深度学习, YOLOv4, 图像识别, GhostNet, 深度可分离卷积模块, 轻量级网络

Abstract:

In view of the power inspection problem caused by various foreign objects on the transmission line, the deep learning image recognition method can be used for detection. This study proposes an improved lightweight network detection algorithm model. By replacing the backbone feature extraction network of YOLOv4 with lightweight neural network GhostNet, the redundancy of feature map generated by image input calculation is reduced. The PANet module of YOLOv4 is modified, and the depth separable convolution module is used to replace the common convolution module, which can reduce the amount of parameter calculation. The results show that, compared with the original YOLOv4 detection algorithm, when the IOU threshold is 0.5, the average accuracy of the improved algorithm decreases by 2.1%, but the detection speed is 2.21 times that of the original algorithm, and the parameter calculation amount is only 17.84% of the original algorithm. The comparison with other algorithms shows that the parameter performance of the proposed algorithm meets the demand. Under the condition of maintaining high accuracy, the detection speed of the proposed algorithm is improved and the computation amount is reduced, which proves the effectiveness and feasibility of the proposed algorithm in target detection.

Key words: transmission line, deep learning, YOLOv4, image recognition, GhostNet, deep separable convolution module, lightweight network

中图分类号:

TP391.4

唐政,张会林,马立新,刘金芝,王昊. 轻量级网络算法对输电线路上异物目标的识别[J]. 电子科技, 2023, 36(4): 71-77.

TANG Zheng,ZHANG Huilin,MA Lixin,LIU Jinzhi,WANG Hao. Identification of Foreign Objects on Transmission Lines Using Lightweight Network Algorithm[J]. Electronic Science and Technology, 2023, 36(4): 71-77.

图/表 12

图1

图2

图3

图4

图5

表1

图6

图7

图8

图9

表2

表3

参考文献 18

[1]	吕明, 罗新, 刘正云. 无人直升机线路巡检技术实用化研究[J]. 湖北电力, 2012, 36(3):10-12.
	Lü Ming, Luo Xin, Liu Zhengyun. Unmanned helicopter study on the application of inspection and detections of power transmission line[J]. Hubei Electric Power, 2012, 36(3):10-12.
[2]	张少平. 输电线路典型目标图像识别技术研究[D]. 南京: 南京航空航天大学, 2012.
	Zhang Shaoping. Research on image recognition for typical objects of transmission line[D]. Nanjing: Nanjing University of Aeronautics and Astronautics, 2012.
[3]	王珂珂. 基于样本生成的输电线路多目标检测算法研究[D]. 阜新: 辽宁工程技术大学, 2020.
	Wang Keke. Research on multi-target detection algorithm of transmission line based on sample generation[D]. Fuxin: Liaoning Technical University, 2020.
[4]	郝帅, 马瑞泽, 赵新生, 等. 基于卷积块注意模型的YOLO-v3输电线路故障检测方法[J]. 电网技术, 2021, 45(8):2979-2987.
	Hao Shuai, Ma Ruize, Zhao Xinsheng, et al. Fault detection of YOLOv3 transmission line based on convolutional block attention model[J]. Power System Technology, 2021, 45(8):2979-2987.
[5]	易继禹, 陈慈发, 龚国强. 基于改进Faster RCNN的输电线路航拍绝缘子检测[J]. 计算机工程, 2021, 47(6):292-298. doi: 10.1063/1.31442
	Yi Jiyu, Chen Cifa, Gong Guoqiang. Aerial photo insulator detection of transmission line based on improved Faster RCNN[J]. Computer Engineering, 2021, 47(6):292-298. doi: 10.1063/1.31442
[6]	颜宏文, 陈金鑫. 基于改进YOLOv3的绝缘子串定位与状态识别方法[J]. 高电压技术, 2020, 46(2):423-432.
	Yan Hongwen, Chen Jinxin. Insulator string positioning and state recognition method based on improved YOLOv3 algorithm[J]. High Voltage Engineering, 2020, 46(2):423-432.
[7]	陈炳煌, 缪希仁, 江灏, 等. 融合粒子群与极限学习机的输电杆塔灾害分类方法[J]. 郑州大学学报(工学版), 2021, 42(4):77-83.
	Chen Binghuang, Miao Xiren, Jiang Hao, et al. A method for disaster status classification of transmission line towers by integrating particle swarm optimization and extreme learning machine[J]. Journal of Zhengzhou University (Engineering Science), 2021, 42(4):77-83.
[8]	吴鹏, 姜海波, 王永强, 等. 参基于图像切片的移动端输电线路鸟类检测算法研究[J]. 计算机与数字工程, 2021, 49(4):846-851.
	Wu Peng, Jiang Haibo, Wang Yongqiang, et al. Research on bird detection algorithm of transmission line based on image slice[J]. Computer and Digital Engineering, 2021, 49(4):846-851.
[9]	Liao X F, Zeng X F. Review of target detection algorithm based on deep learning[C]. Chongqing: Proceedings of the International Conference on Artificial Intelligence and Communication Technology, 2020.
[10]	Zhao Z M, Lei X Y. Improved real-time pedestrian detection method[C]. Dalian: IEEE the Seventh International Conference on Computer Science and Network Technology, 2019.
[11]	Chen L Y, Zheng M C, Duan S Q, et al. Underwater target recognition based on improved YOLOv4 neural network[J]. Electronics, 2021, 14(10):1634-1634.
[12]	Wei B Y, Shen X L. Remote sensing scene classification based on improved GhostNet[C]. Wuhan: Proceedings of the International Conference on Computer Science and Communication Technology, 2020.
[13]	肖红, 张瑶瑶, 张福禄. 改进的卷积神经网络及在地层识别中的应用[J]. 计算机技术与发展, 2021, 31(9):167-172.
	Xiao Hong, Zhang Yaoyao, Zhang Fulu. Improved convolutional neural network and its application in stratigraphic identification[J]. Computer Technology and Development, 2021, 31(9):167-172.
[14]	温博阁. 基于深度可分离卷积的多目标追踪神经网络研究[J]. 大连交通大学学报, 2021, 42(5):111-114.
	Wen Boge. Multi-target tracking neural network based on depthwise separable convolutions[J]. Journal of Dalian Jiaotong University, 2021, 42 (5):111-114.
[15]	薛钰洁. 基于多种数据库的改进YOLO算法研究[D]. 大连: 大连理工大学, 2021.
	Xue Yujie. Research on improved YOLO algorithm based on multiple databases[D]. Dalian: Dalian University of Technology, 2021.
[16]	Qing Y H, Liu W Y, Feng L Y, et al. Improved YOLO network for free-angle remote sensing target detection[J]. Remote Sensing, 2021, 11(13):2171-2171. doi: 10.3390/rs11182171
[17]	张莹, 刘子龙, 万伟. 基于Faster R-CNN的无人机车辆目标检测[J]. 电子科技, 2021, 34(11):11-20.
	Zhang Ying, Liu Zilong, Wan Wei. UAV vehicle target detection based on Faster R-CNN[J]. Electronic Science and Technology, 2021, 34(11):11-20.
[18]	邵叶秦, 周昆阳, 郑泽斌, 等. 基于改进的轻量级YOLOv3的交通信号灯检测与识别[J]. 南通大学学报(自然科学版), 2021, 20(3):34-40.
	Shao Yeqin, Zhou Kunyang, Zheng Zebin, et al. Traffic light detection and recognition based on improved lightweight YOLOv3[J]. Journal of Nantong University(Natural Science Edition), 2021, 20(3):34-40.

参数名称	参数值
输入尺寸	608×608
训练次数	100
训练集与测试集比例划分	9∶1
Batch_size	12
初始学习率	0.001

GhostNet	深度可分离卷积	mAP/%	FPS/frames·s^-1	FLOPs
		98.04	12.05	64 363 101
√		96.35	16.32	39 689 409
	√	94.12	17.12	36 102 237
√	√	95.94	26.67	11 428 545

算法模型	mAP/%	FPS/ frames·s^-1	FLOPs
本文算法	95.94	26.67	11 428 545
YOLOv4	98.04	12.05	64 363 101
YOLOv3	68.59	25.31	61 949 149
YOLOv4-tiny	73.35	72.31	6 102 237
文献[17]	88.82	6.21	137 078 239
文献[18]	83.76	24.03	8 700 432