基于EfficientDet的轨道扣件识别与检测
Track fastener identification and detection based on EfficientDet
邹文武 1许贵阳 1白堂博1
作者信息
- 1. 北京建筑大学机电与车辆工程学院,北京 100044;北京建筑大学城市轨道交通车辆服役性能保障北京市重点实验室,北京 100044
- 折叠
摘要
扣件作为轨道线路的重要部件,其缺陷状态复杂多变,当前在实际运维中主要采用人工巡检的方式检测,该方式效率低、耗时长,检测结果依赖人员的熟练程度.针对以上问题,提出一种基于改进EfficientDet的扣件状态检测方法.首先,对图像数据集进行标注与数据增强;然后利用信道修剪算法对EfficientDet网络进行优化、训练与识别;最后进行实际轨道线路图像采集实验,并与YOLOv3(you only look once v3)和Faster R-CNN(faster region-based convolutional neural network)进行对比,选择合适的评价模型和标准进行结果分析.结果表明:所提出的改进方法对铁路缺陷扣件的检测准确率达到96.82％,检测效率和检测精度较其他2种方法有明显提高,且参数量是YOLOv3的1/5,表明其在目标检测应用中具有很高的潜力.
Abstract
As an important part of track lines,the defect status of fasteners is complex and changeable.Manual inspection methods are currently mainly used in actual operation and maintenance,which are inefficient and time-consuming,and the detection results depend on the proficiency of inspection personnel.To solve the above problems,a method for detecting the status of fasteners based on improved EfficientDet is proposed.Firstly,the image data set is marked and augmented.Then,the channel pruning algorithm is used to optimize,train and identify the EfficientDet network.Finally,the actual track line image acquisition experiment is carried out and compared with YOLOv3(you only look once v3)and Faster R-CNN(faster region-based convolutional neural network),and the appropriate evaluation model and standard are selected for result analysis.The results show that the improved method proposed in this paper has a detection accuracy of 94.4％for railway defective fasteners,the detection efficiency and detection accuracy are significantly improved compared with the other two methods,and the parameter quantity of the proposed method is 1/5 of YOLOv3,which indicating that the proposed method has extremely high potential in target detection applications.
关键词
EfficientDet/扣件检测/通道修剪/数据增强/深度学习Key words
EfficientDet/fastener detection/channel trimming/data augmentation/deep learning引用本文复制引用
基金项目
国家自然科学基金面上项目(51975038)
北京市自然科学基金重点项目(KZ202010016025)
出版年
2024
NETL
NSTL
万方数据