深度学习在输电线路无人机巡检中识别防振锤缺陷的应用

Application of deep learning in vibration damper defect identification to UAV inspection of transmission line

基于深度学习技术,对无人机巡检架空线路防振锤缺陷识别的应用开展了研究。针对巡检图像背景复杂,防振锤识别目标小、易出现漏检的特点,在更快的基于区域的卷积神经网络(faster region-based convolution neural network, Faster R-CNN)算法基础上,利用光学矫正和综合去噪对图像进行预处理,用深度残差网络(deep residual network, ResNet)提取特征图,增大、增多训练尺度对算法进行优化。通过Gabor滤波器和分析红绿蓝(red green blue, RGB)颜色通道特征分别对防振锤损坏和锈蚀缺陷进行识别。实验结果表明:该方法对防振锤缺陷有较好的识别效果,召回率达到92.56%,精确度达到98.24%,优于现有的其他目标检测方法。

The application of unmanned aerial vehicle(UAV) inspection of vibration damper defect recognition of overhead line using deep learning is studied in this paper. In view of the inspection images of the complex background, the small target of damper recognition and easy to miss detection, this paper adopts a series of pretreatment techniques. Based on the faster region-based convolution neural network(Faster R-CNN)algorithm, the image is preprocessed by optical correction and comprehensive de-noising, the deep residual network(ResNet) is used to extract the feature map, and the training scale is increased to optimize the algorithm. Gabor filter and red-green-blue(RGB) color feature analysis are used to identify the damage and corrosion defects of damper. According to the experiment results, the method is very efficient on detecting the damper defects. With the recall rate of 92.56%, and the accuracy of 98.24%, the method is better than other existing target detection methods.