摘要
提出一种基于改进EfficientDet和YOLOv7算法的钢轨损伤检测方法,以应对传统人工巡检效率低和精度不足的问题。该方法结合了EfficientNet与双向特征金字塔网络(BiFPN),显著提高了多尺度特征的处理能力,并通过引入更深的卷积层增强了特征提取能力。通过使用高清相机和光照补偿装置对钢轨表面进行图像采集,利用卷积神经网络对图像进行处理分析,实现快速、精准的钢轨损伤检测。实验结果表明,相比于YOLOv7、YOLOv5和YOLOv3,EfficientDet-YOLOv7算法在平均精度均值(mAP)、查准率(Precision)和查全率(Recall)等指标上分别提升了9.27%,8.50%和9.20%,并且具有更高的计算效率和收敛速度。
This paper proposed a rail damage detection method based on an improved EfficientDet and YOLOv7 algorithm to address the issues of low efficiency and insufficient accuracy in traditional manual inspections.The method combined EfficientNet with the bi-directional feature pyramid network(BiFPN),significantly enhancing the processing ability of multi-scale features.Additionally,deeper convolutional layers were introduced to strengthen feature extraction capabilities.Using high-definition cameras and lighting compensation devices to capture images of the rail surface,the convolutional neural network processed and analyzed the images to achieve rapid and accurate rail damage detection.Experimental results show that,compared to YOLOv7,YOLOv5,and YOLOv3,the EfficientDet-YOLOv7 algorithm improves the mean average precision(mAP),precision,and recall by 9.27%,8.50%,and 9.20%,respectively,while also offering higher computational efficiency and faster convergence speed.
作者
闫龙
袁花明
汤超
YAN Long;YUAN Huaming;TANG Chao(Shaanxi Jingshen Railway Co..Ltd,Yulin 719000,Shaanxi,China;Transport Command and Information Technology Research Institute,CRSC Research&Design Institute Group Co.,Ltd.,Beijing 100000,China)
出处
《铁路物流》
2025年第8期62-67,共6页
Railway Logisitics
作者简介
通信作者:汤超(1980-),男,北京人,北京全路通信信号研究设计院集团有限公司运输指挥与信息技术研究院工程师。
