基于YOLOv5的电力线和杆塔实时检测算法研究
2022,30(11):77-84
摘要：针对当前电力线路检测中存在深度学习网络参数量大、计算复杂度高等问题；在YOLOv5的基础上提出一种电力线和杆塔的实时检测算法；通过减少Bottleneck数量来简化特征提取层网络结构，使用深度可分离卷积技术实现模型计算量的降低；分析电力线目标框筛选机制，改进(Non-Maximum Suppression)NMS算法，提升模型目标检测精度；实验结果表明，对Bottleneck的改进在识别精度有所提高的情况下能有效降低模型的参数量，模型检测准确率和召回率分别达到94%与95%，体积压缩了20.7%，在Jetson Nano嵌入式平台上检测速度达到17.2 fps，对两类电力线路目标检测达到较高的识别率和实时性，对无人机电力巡检导航有较好的参考价值。
关键词：电力线路检测;Bottleneck;深度可分离卷积;NMS;嵌入式平台
Research on Real-time Detection Algorithm of Power Line and Pole Tower Based on YOLOv5
Abstract：Aiming at the problems of large number of parameters and high computational complexity of deep learning network in current power line detection; Based on YOLOv5, a real-time detection algorithm for power lines and towers is proposed. The network structure of feature extraction layer is simplified by reducing the number of Bottleneck, and the depth-separable convolution technique is used to reduce the computational amount of model. The mechanism of power line target box screening was analyzed, and the (non-maximum Suppression) NMS algorithmwas improved to improve the model target detection accuracy. The experimental results show that the improvement of Bottleneck can effectively reduce the number of parameters of models when the recognition accuracy increases. The model detection accuracy and recall rate reach 94% and 95%, respectively, and the volume is compressed by 20.7%. The detection speed on Jetson Nano embedded platform reaches 17.2 FPS. The detection of two kinds of power line targets achieves high recognition rate and real-time performance, which has a good reference value for UAV power inspection and navigation.
Key words：Transmission line detection; Bottleneck; Depth separable convolution; NMS; Embedded platform
收稿日期：2022-08-08
基金项目：
