基于IR-DETR的轨道图像增强及扣件损伤检测方法  

Track Image Enhancement and Fastener Damage Detection Method Based on IR⁃DETR

作　　者：段嘉明 白堂博 许贵阳 宗浩 付浩辰 DUAN Jiaming;BAI Tangbo;XU Guiyang;ZONG Hao;FU Haochen(School of Mechanical,Electrical and Vehicle Engineering,Beijing University of Civil Engineering and Architecture,Beijing 100044,China;Beijing Key Laboratory of Performance Guarantee on Urban Rail Transit Vehicles,Beijing 100044,China)

机构地区：[1]北京建筑大学机电与车辆工程学院,北京100044 [2]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

出　　处：《铁道建筑》2025年第1期49-53,共5页Railway Engineering

基　　金：国家自然科学基金(52272385);北京市自然科学基金(L211007)。

摘　　要：为解决在轨道扣件检测中常见的光线不足导致的图像模糊昏暗、扣件特征提取不足等问题,提出一种基于InstructIR(Instruct Image Restoration)与RT-DETR(Real Time Detection Transformer)的综合方法IR-DETR。在图像预处理方面,针对轨道检测图像的特点,提出基于InstructIR的图像增强方法,根据智能检测的需求增强图像特征。在扣件损伤检测方面,优化RT-DETR模型,引入可学习位置编码(Learned Positional Encoding,LPE),用于对序列中的位置信息进行编码,并在主干部分融合可变形卷积DCNv2(Deformable ConvNets v2),进一步提升模型的感知能力与特征表达能力。用优化前后的数据集作为输入,对IR-DETR及主流模型进行了对比试验。结果表明:改进后的模型平均检测精度提高了2.1%,在参数量基本不变的情况下检测速度提高了18.6%。To solve the common problems of image blurring and insufficient feature extraction caused by insufficient lighting in track fastener detection,a comprehensive method IR-DETR based on Instruction Image Restoration(InstructIR)and Real-time Detection Transformer(RT-DETR)was proposed.In terms of image preprocessing,an image enhancement method based on InstructIR was proposed to address the characteristics of track detection images and enhance image features according to the requirements of intelligent detection.In terms of fastener damage detection,the RT-DETR model was optimized by introducing Learned Positional Encoding(LPE)to encode position information in the sequence,and fusing Deformable ConvNetsv2(DCNv2)in the backbone to further enhance the model’s perception and feature expression capabilities.Comparative experiments were conducted on IR-DETR and mainstream models,and pre and post optimization datasets were used as inputs.The results showed that the improved model increases the average detection accuracy by 2.1%,and the detection speed is increased by 18.6%while keeping the number of parameters basically unchanged.

关 键 词：轨道 实时检测 试验研究 扣件伤损 图像增强 智能检测 

分 类 号：U213.4[交通运输工程—道路与铁道工程]