摘要
依靠铁路工人人工巡检钢轨表面缺陷,存在较大误差,耗费大量人力物力,制约了我国铁路行业的健康发展。针对上述问题,论文设计了一种基于CNN-Transformer的钢轨表面缺陷识别方法;使用Transformer层的堆叠代替标准卷积的矩阵乘法用以对卷积提取的高层语义信息进行全局建模;同时引入轻量级的CNN网络GhostNet,提取图像特征,以减少计算参数,补偿因使用Transformer而缺乏归纳偏置的缺点。结果显示,基于论文方法的钢轨表面缺陷识别精度达到94.51%,高于VGG16、ResNet50、MobileNet等传统的CNN网络,且计算成本更低,为机器视觉在钢轨维护领域的应用提供了重要参考。
Relying on the manual inspection of rail surface defects by railway workers,there is a large error,which consumes a lot of manpower and material resources,and restricts the healthy development of China's railway industry.To solve these prob⁃lems,this paper designs a rail surface defect recognition method based on CNN Transformer.The stack of Transformer layer is used to replace the matrix multiplication of standard convolution for global modeling of high-level semantic information extracted by con⁃volution.At the same time,the lightweight CNN network GhostNet is introduced to extract image features to reduce calculation pa⁃rameters and compensate for the lack of inductive bias due to the use of Transformer.The results show that the rail surface defect rec⁃ognition accuracy based on this method reaches 94.51%,higher than VGG16,ResNet50,MobileNet and other traditional CNN net⁃works,and the calculation cost is lower,which provides an important reference for the application of machine vision in the field of rail maintenance.
作者
张春光
许嘉瑞
马骏
ZHANG Chunguang;XU Jiarui;MA Jun(School of Automation and Electrical Engineering,Dalian Jiaotong University,Dalian 116028;School of Mechanical Engineering,Dalian Jiaotong University,Dalian 116028)
出处
《计算机与数字工程》
2025年第2期540-544,共5页
Computer & Digital Engineering
