作者:宋立博 ，费燕琼
\n

Authors:SONG Li-bo ,FEI Yan-qiong
摘要:摘要:针对电力线快速智能巡检需求，提出在飞行器上安装边缘设备进行智能检测的设想。为提高检测性能，依据原始YOLOv4-tiny深度网络结构设计了以Resblock-D轻量化网络为特征提取网络和Resblock-D+CSPDarknet53-tiny为主干的新型YOLOv4-tiny算法和网络。在Jetson NANO上在搭建成功GPU版Darknet深度框架用于训练、部署和测试。以ImageNet标准图像库中的power line图片集及街拍电力线图片为基础，建立Pascal VOC格式的绝缘子数据集,训练成功无预训练权重文件的绝缘子识别网络。选择平均均值误差(mAP)最高的权重文件，在Jetson NANO上进行绝缘子智能识别实验。与标准YOLOv4-tiny算法相比，新型YOLOv4-tiny算法权重文件和计算量仅为58%和66.7%，检出速度快约16%，检出数量高约10%，误检数量低10%,证实其高效性和实用性。
Abstract:Abstract:Conforming to the rapid increasing requirements of fast and intelligent inspection of power lines, the idea of installing edge device on aircraft for intelligent inspection is put forward. The Resblock-D lightweight network is selected as the feature extraction network, and the new YOLOv4-tiny algorithm and deep network based on the standard YOLOv4-tiny structure is designed as the Resblock-D and the CSPDarknet53-tiny are used as the main backbone.Then, the related GPU version Darknet deep network frame was built on Jetson NANO for training, deploying and testing.On basis of the power line images sets from ImageNet standard library and the street shot power line pictures, the isolator data set is established accordingly in the format of standard Pascal VOC.Under the Darknet deep network frame, the detector train command is utilized to train the new YOLOv4-tiny network without the pre-training weight file successfully. Selecting the weight file of the highest mAP(mean average precision), the isolator intelligent identification experiment was carried out on Jetson NANO. In term of 58% weight files, 66.7% computation, higher 10% detection quantity and lower 10% false detection quantity, it is proved that the new YOLOv4-tiny algorithm is efficient and practical in comparison to the standard YOLOv4-tiny algorithm.

PDF全文下载地址:

可免费Download/下载PDF全文