基于深度视觉算法的轨面伤损检测方法  

Rail surface damage detection method based on depth vision algorithm

作　　者：王保成[1] 袁昊 韩峰[1] 王超 李佳恒 WANG Baocheng;YUAN Hao;HAN Feng;WANG Chao;LI Jiaheng(School of Civil Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学土木工程学院,甘肃兰州730070

出　　处：《实验技术与管理》2024年第9期84-91,共8页Experimental Technology and Management

基　　金：中国国家铁路集团有限公司重点课题(N2023X050);兰州交通大学重点研发项目(LZJTU-ZDYF2305)。

摘　　要：针对现有目标检测器存在的推理延迟、不稳定和高计算成本等问题,提出一种基于深度学习理论的创新算法RT-DETR(RT-DETR-L),实现了对钢轨表面伤损的高效精细化检测。基于该算法设计的目标检测实验方案,去除了传统目标检测算法中的非极大值抑制(NMS)后处理步骤;引入了一个解耦单尺度内部交互和跨尺度融合的高效混合编码器;提出了一种IoU-aware初始化对象查询机制,并重新定义了目标函数。实验结果表明,该方案能有效提高算法在检测钢轨表面伤损时的准确率和召回率,在检测剥离掉块、疲劳裂纹、接头方面表现出色,准确率分别为95.1%、93.8%和99.5%,检测速度为8.62 ms/帧,参数量仅为4.2 M。该研究成果能够为钢轨养护维修提供一种准确高效的检测方案。[Objective]The expansion of railroad operations has facilitated the research on the timely and accurate detection of rail surface injuries.Traditional methods of rail surface defect detection heavily relied on manual inspections.Although these methods offer benefits such as simple operation and low cost,they possess drawbacks such as low efficiency,high rates of missed defects,and poor real-time performance.Automated detection techniques such as ultrasonic,eddy current,and magnetic flux leakage detection are significant in rail damage detection.However,they possess limitations such as high dependence on hardware and the need for skylight point operation,leaving some rail injuries unaddressed in a timely manner.Furthermore,despite the advances in existing deep learning-based algorithms for rail defect detection,they face issues such as inference delays,instability,and high computational costs,failing to meet the real-time requirements of field operations.To overcome these limitations,we propose an innovative algorithm based on deep learning theory,referred to as RT-DETR,to achieve efficient and precise detection of rail surface injuries.[Methods]The RT-DETR algorithm eliminates the need for non-maximum suppression(NMS)in traditional target detection algorithms,which helps avoid the inference delays and instability associated with existing detectors.Based on the single-scale feature interaction module(AIFI)with the attention mechanism and the cross-scale feature fusion module(CCFM)based on CNN,we introduce an efficient hybrid encoder that decouples single-scale internal interactions and cross-scale fusion,replacing the traditional transformer encoder.An IoU-aware initialized object query mechanism is proposed to enhance the initialization of decoder queries.The objective function is redefined to better balance classification scores with IOU scores and improve detection performance.This further ensures consistency in the classification and localization of positive-sample targets,thus enabling the model to produce hi

关 键 词：钢轨表面 伤损检测 NMS 混合编码器 loU-aware 

分 类 号：U216.3[交通运输工程—道路与铁道工程]