Abstract: As an important target in power inspection, insulator detection is highly valued. In view of the low efficiency and poor robustness of traditional UAV image insulator detection algorithm which needs manual feature extraction, an improved YOLOv3 insulator detection algorithm is proposed based on the analysis of insulator image data set, combined with YOLOv3 target detection algorithm and Inception-Resnet-v2 classification algorithm, by changing the network structure to increase the network width, improving the loss function according to the characteristics of the data set, and using [k]-means algorithm to select anchor box, YOLOv3 is improved. For 52 insulator images of test set, the total detection time of YOLOv3 model is 5.12?s, the improved YOLOv3 takes 5.48?s, the AP value of YOLOv3 is 69.69%, and the improved AP value is 71.93%. The experimental results show that compared with the original YOLOv3 algorithm, the improved YOLOv3 algorithm can better adapt to the data characteristics of insulator image data set captured by UAV, and ensure the same detection speed. The AP is increased by 2.24 pecentage points.

Key words: neural network, insulator, YOLOv3, Inception-Resnet

摘要： 绝缘子作为电力巡检中的重要目标，对其检测受到高度重视。针对传统无人机拍摄图像绝缘子检测算法需要通过人工提取特征进行检测，效率低且鲁棒性差等问题，通过对绝缘子图像数据集进行分析，结合YOLOv3目标检测算法与Inception-Resnet-v2分类算法，提出一种改进的YOLOv3绝缘子检测算法，该方法分别从增加Inception-Resnet模块，通过改变网络结构以增加网络宽度，根据数据集特性改进损失函数以及利用[k]-means算法进行锚点框选择三个方面对YOLOv3进行改进。针对测试集52张绝缘子图像，YOLOv3模型检测总耗时5.12?s，改进后的YOLOv3耗时5.48?s，YOLOv3的绝缘子AP值为69.69%，改进后AP值为71.93%，实验结果表明，相较于原始YOLOv3算法，改进后的YOLOv3算法能够更好地适应无人机拍摄的绝缘子图像数据集的数据特征，在保障检测速度的同时AP提高了2.24个百分点。

关键词: 神经网络, 绝缘子, YOLOv3, Inception-Resnet