基于串联Faster R-CNN的绝缘子自爆检测识别

Detection and Identification of Insulator Self-explosion Based on Cascaded Faster R-CNN Network

高压输电线路中绝缘子自爆造成的缺损会严重危害输电线路的安全运行。针对无人机巡检过程中获取的图像光线明暗不定、背景复杂、小目标等因素导致的绝缘子检测难度大、自爆识别准确率低等问题,提出了一种基于串联Faster R-CNN网络的无人机图像中绝缘子检测和自爆识别的算法。该算法分为两个阶段,分别串联使用深度学习中具有强大目标检测能力的Faster R-CNN网络实现对无人机高压输电线路图像中绝缘子自爆的检测和识别。第一阶段使用Faster R-CNN网络检测出无人机高压输电线路图像中绝缘子,第二阶段使用Faster R-CNN网络对检测出的绝缘子图像进行自爆识别。Faster R-CNN网络的目标特征提取采用“特征金字塔网络”(feature pyramid network,FPN),结合区域建议网络(region proposal network,RPN)生成候选区域,提高了对绝缘子检测和自爆识别的性能。通过对无人机图像进行绝缘子检测和常见自爆识别的实验,结果表明,相比于直接使用Faster R-CNN以及SVM、VGG等对绝缘子自爆进行检测和识别,该算法提高了绝缘子自爆检测和识别的精度。

Defects caused by self-explosion of insulators in high voltage transmission lines can seriously endanger the safe operation of transmission lines.An algorithm based on series Faster R-CNN network is proposed for detecting and identifying the insulators in unmanned aerial vehicle images,which is difficult to detect and has low accuracy due to the unstable light,complex background,small targets and other factors.The algorithm is divided into two phases,using Faster R-CNN network with strong target detection capability in deep learning to detect and identify insulator self-explosion in unmanned aerial vehicle high voltage transmission line images.The first phase uses the Faster R-CNN network to detect the insulators in the high voltage transmission line image of the UAV,and the second phase uses the Faster R-CNN network to identify the self-explosion of the detected insulators image.Target feature extraction of Faster R-CNN network uses feature pyramid network(FPN)to generate candidate regions in combination with region proposal network(RPN),which improves the performance of insulator detection and self-explosion recognition.Through the experiments of detecting and identifying the common self-explosion of the insulator in the unmanned aerial vehicle image,it is showed that the proposed algorithm improves the accuracy of detecting and identifying the self-explosion of the insulator compared with using Faster R-CNN,SVM and VGG directly.