基于图像增广与迁移学习的输电线路金具多目标实时检测方法

Multi-objective real-time detection method of transmission line fittings based on image augmentation and transfer learning

架空输电线路金具的状态评估工作对于线路的可靠运行至关重要,金具的检测是评估工作的重要一环。针对金具识别检测中数据集人工标注的工作量大,以及难以兼顾高精度和快速性问题,提出一种基于YOLOX网络改进的输电线路金具检测方法。将无人机拍摄的金具图像进行增广预处理丰富数据集,骨干网络采用在线Mosaic、Mixup增强方式,引入基于特征提取的迁移学习并采用余弦退火学习率进行两阶段模型训练。实验结果表明,改进后的方法对各类金具检测的平均精度均值提高了18.32%,与Faster R-CNN等5种主流检测模型相比,所提方法平均检测精度均值最高,且检测速度仅次于YOLOv3,能够更加快速、精准地识别各类金具,并在一定程度上减少人工标注的工作量。

The state assessment of overhead transmission line fittings is crucial to the reliable operation of the line, and the detection of the fittings is an important part of the assessment work. In response to the heavy workload of manual labeling of datasets in identification and detection of fittings, as well as the difficulty of balancing high precision and rapidity, an improved transmission line fittings detection method based on YOLOX network is proposed. The fitting images captured by UAV are augmented with preprocessing to enrich the datasets. The backbone network adopts the enhancement methods of online Mosaic and Mixup. The transfer learning based on feature extraction is introduced and the cosine annealing learning rate is used for two-stage model training. The experimental results show that the mean average precision of the improved method for the detection of all types of fittings is improved by 18.32%. Compared with five mainstream detection models such as Faster R-CNN algorithm, the mean average precision of proposed method is the highest, and its detection speed is lower than YOLOv3’s, which can identify various types of fittings more quickly and accurately, and reduce the workload of manual labeling to a certain extent.