基于改进YOLOV5n的绝缘子和间隔棒检测算法

An Insulator and Spacer Detection Algorithm Based on the Improved YOLOV5n

轻量级神经网络的出现显著降低了目标检测算法在移动端部署的难度,当前已有许多运算量小、精度较高的卷积神经网络在多个公共数据集上取得了不错的效果。然而,在基于图像的电力巡检领域,图像中目标检测与识别的效率对于及时排除电力故障具有重要意义,尤其是针对基于无人机的巡检,实现在线实时的故障检测更有意义。为了实现绝缘子和间隔棒这些重要巡检目标的移动端实时检测,本文提出了一种基于YOLOV5n的针对电力设备检测与分类的轻量级网络模型,算法在YOLOV5n的基础上优化网络,通过减少一系列的卷积层并舍去一部分的捷径分支,提高网络的并行程度并降低网络的深度。最终设计出模型更轻量、精确度更高的YOLOV5n-1、YOLOV5n-2,基于自建的电力巡检数据集进行测试,实验结果表明,提出的算法比YOLOV5n减少了27%的GFLOPs,检测时间降低了24%,降低了硬件要求,更适合在移动端部署。

The emergence of lightweight neural networks has significantly reduced the difficulty of deploying object detection algorithms on mobile devices.Currently,many convolutional neural networks with low computational complexity and high accuracy have achieved good results on multiple common datasets.However,in the field of image-based power inspection,the efficiency of object detection and recognition in images is of great significance for timely troubleshooting of power faults.Especially for UAV-based inspections,it is more meaningful to realize online real-time fault detection.In order to achieve real-time detection of important inspection targets such as insulators and spacers on mobile terminals,a lightweight network model based on YOLOV5n for power equipment detection and classification is proposed in this paper.The algorithm optimizes the network based on YOLOV5n by reducing a series of convolutional layers and discarding some shortcut branches,which improves the parallelism of the network and reduces its depth.Finally,YOLOV5n-1 and YOLOV5n-2 are designed with lighter and more accurate models.Tests have been conducted based on the self-built power inspection dataset,and the experimental results show that the proposed algorithm,as compared with YOLOV5n,reduces the GFLOPs by 27%,reduces the detection time by 24%,reduces hardware requirements,which is more suitable for deployment on mobile devices.