基于改进YOLOv4输电线关键部件实时检测方法

针对输电线路维护过程中的典型缺陷识别问题,为提高无人机(unmanned aerial vehicle, UAV)自主巡检的智能化程度,提出基于改进YOLOv4的无人机输电线关键部件实时检测模型。根据无人机视角下输电线典型目标的特点,结合MobileNet重新设计了一种轻量的特征提取网络来获取更高的特征提取效率,利用空洞模块增强感受野减少小目标的信息损失;在特征融合模块中添加自适应路径融合网络来融合更多的位置信息和语义信息,提高了多尺度目标的检测精度,减少了目标的误报率。采用构建的无人机输电线关键部件数据集来评估提出的模型。结果表明:基于YOLOv4改进的网络能够在无人机机载端实现实时多尺度目标检测,模型的平均准确率可达到92.76%,检测速度可达到32帧/秒,能够满足无人机嵌入式平台上实时检测的需求。

Real-time Inspection Model for Key Components of Transmission Lines Based on Improved YOLOv4

In order to address the typical defect recognition problem in transmission line maintenance process and improve the intelligence of UAV autonomous inspection, a real-time detection model of UAV transmission line key components based on improved YOLOv4 was proposed. First, according to the characteristics of typical targets of transmission lines from UAV''s viewpoint, a lightweight feature extraction network was redesigned with MobileNet to obtain higher feature extraction efficiency, and the cavity module was used to enhance the perceptual field to reduce the information loss of small objects; an adaptive path fusion network was added to the feature fusion module to fuse more location and semantic information, which improved the detection of multi-scale targets accuracy and reduce the false alarm rate of targets. The results show that the improved network based on YOLOv4 can achieve real-time multiscale detection on the UAV airborne side, and the mAP(mean average precision) achieved 92.76% and the detection speed reached 32 frames/second, which meet the demand of real-time detection on the UAV embedded platform.