基于卷积神经网络和航拍图像的电力线路绝缘子缺陷识别

Insulator Defect Recognition of Power Line Based on Convolution Neural Network and Aerial Images

由于电力线路绝缘子故障导致输电系统故障,基于高空作业平台的绝缘子检测系统得到了广泛的应用。绝缘子缺陷检测是在航空图像复杂背景下进行的,检测识别具有一定的挑战性。而基于主观的特征或浅层学习技术的传统方法只能在特定的检测条件下定位绝缘子并检测故障。为此,本研究讨论了利用航空图像自动检测绝缘子缺陷,精确定位从实际检测环境采集的输入图像中出现的绝缘子缺陷。提出了一种新颖的深卷积神经网络(convolutional neural network, CNN)级联结构,用于绝缘子缺陷的定位和检测。级联网络使用基于区域建议网络的CNN将缺陷检测转化为两级目标检测问题。针对实际检测环境中缺陷图像的稀缺性,提出了一种数据增强方法,该方法包括四种操作:1)仿射变换;2)绝缘子分割和背景融合;3)高斯模糊;4)亮度变换。有实验结果可见,该方法的缺陷检测精度和召回率分别为0.91和0.96,均优于现阶段常用的方案,能够满足绝缘子缺陷检测的鲁棒性和准确性要求。

Due to the transmission system failure caused by power line insulator fault, the insulator detection system based on aerial work platform has been widely used. Insulator defect detection is carried out in the complex background of aerial images, and detection and recognition has certain challenges. However, traditional methods based on subjective features or shallow learning technology can only locate insulators and detect faults under specific detection conditions. Therefore, the author discusses the automatic detection of insulator defects using aerial images, and accurately locate the insulator defects in the input images collected from the actual detection environment. A novel cascaded structure of deep convolutional neural network(CNN) is proposed for insulator defect location and detection. In cascaded network, CNN based on area recommendation network is used to transform defect detection into two-level target detection problem. In view of the scarcity of defect image in actual detection environment, a data enhancement method is proposed, which includes four operations: 1) affine transformation;2) insulator segmentation and background fusion;3) Gaussian blur;4) brightness transformation. Experimental results show that the defect detection accuracy and recall rate of this method are 0.91 and 0.96 respectively, which are better than the commonly used schemes at present stage, and can meet the requirements of robustness and accuracy of insulator defect detection.