一种少样本条件下的钢轨伤损图像缺陷检测方法  

A Method for Detecting Defects in Rail Damage Images with Few Samples

作　　者：曾树华 黄银秀[3] 黄昌兵[1,2] ZENG Shuhua;HUANG Yinxiu;HUANG Changbing(Hunan High Speed Railway Operation Safety Assurance Engineering Technology Research Center,Zhuzhou 412006,China;Hunan Vocational College of Railway Technology,Zhuzhou 412006,China;Hunan Chemical Vocational Technology College,Zhuzhou 412006,China)

机构地区：[1]湖南省高铁运行安全保障工程技术研究中心,湖南株洲412006 [2]湖南铁路科技职业技术学院,湖南株洲412006 [3]湖南化工职业技术学院,湖南株洲412006

出　　处：《现代信息科技》2023年第19期134-137,共4页Modern Information Technology

基　　金：湖南省自然科学基金(2020JJ7054)。

摘　　要：为解决钢轨表面伤损检测问题,提出一种少样本条件下的钢轨表面伤损检测方法。首先,设计样本随机组合策略,扩充钢轨表面伤损数据集规模;其次,引入迁移学习方法,在公开大规模数据集上进行迁移学习训练,以获得迁移学习能力,降低对钢轨表面伤损样本的需求数量;最后,加入通道自注意力机制,提高模型的训练速度。实验证明,该方法可有效提高钢轨表面伤损的识别精度。To solve the problem of rail surface damage detection,a rail surface damage detection method with few samples is proposed.Firstly,design a sample random combination strategy to expand the scale of the rail surface damage dataset;secondly,introduce transfer learning methods and conduct transfer learning training on publicly available large-scale datasets to obtain transfer learning capabilities and reduce the demand amount for rail surface damage samples;finally,a channel self attention mechanism is added to improve the training speed of the model.Experiments have shown that this method can effectively improve the recognition accuracy of rail surface damage.

关 键 词：少样本 钢轨表面伤损 迁移学习 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP278[自动化与计算机技术—计算机科学与技术]