基于YOLO-V5的输电线路可视化中视觉分析关键技术的研究及应用

Research and Application of Key Technologies for Visual Analysis of Transmission Lines Based on YOLO-V5

为解决输电线路巡检中以往单一的人工巡线模式和传统视频监控模式实时性、准确性差的问题,本文探索了一种基于深度学习框架YOLO-V5的输电线路隐患目标检测算法,该算法能够对输电线路存在的各类隐患进行较为准确的预警。本研究在对17386张包含各类输电线路隐患的图像进行细致化标注与图像扩增后,使用YOLO-V5在COCO数据集上的预训练模型进行训练,并经过两次学习率的调整后得到最终收敛的模型。该模型在对包含6933张各类输电线路隐患图像和无隐患图像的测试集进行分析后,最终预测结果表明综合准确率为79.10%,综合召回率为73.45%,漏检率为3.93%,误检率为4.97%。本文所提出的隐患识别模型可以对输电线路隐患进行实时识别,识别准确率较高,误报、漏报率较低,能够有效提升输电线路运检效率。

In order to solve the problem of poor real-time and accuracy of single manual patrol mode and traditional video monitoring mode in transmission line patrol inspection, this paper explores a target detection algorithm for transmission line hidden danger based on deep learning framework YOLO-V5, which can provide more accurate warning for all kinds of hidden dangers of transmission line. This study used YOLO-V5 pre-training model on COCO dataset after detailed labeling and image enlargement of 17386 images containing all kinds of potential transmission line hazards. After two adjustments of learning rate, the final convergence model was obtained. After analyzing the test set containing 6933 hidden and non-hidden images, the final prediction results show that the overall accuracy is 79.10%, the comprehensive recall rate is 73.45%, the miss rate is 3.93%, and the false detection rate is 4.97%. The hidden danger identification model proposed in this paper can be used to identify hidden dangers of transmission lines in real-time. It has a high recognition accuracy and a low rate of false positives and false positives, which can effectively improve the efficiency of transmission line inspection.