摘要
铁轨缺陷检测对铁路安全和降低维护成本非常重要。面对铁路网络扩张和缺陷样本稀缺带来的挑战,尤其是小样本条件下的过拟合问题,本文提出了一种基于多尺度WideResNet(Wide Residual Network)的小样本铁轨表面缺陷检测算法,通过数据增强等图像处理技术扩大有限的训练集,提高模型的泛化能力;利用迁移学习策略,通过预训练的WideResNet模型提取多尺度特征,并将其用于铁轨缺陷检测任务,减少对大量标注数据的依赖,加快小样本模型训练快速收敛;设计小样本深度学习模型策略,构建度量学习模块,从有限的标注数据中快速学习并进行有效泛化。结果表明,该算法在10-shot小样本条件下能够有效地检测铁轨表面缺陷,模型平均精度达到83.6%,召回率高达93.8%。
Rail defect detection is crucial for railway safety and reducing maintenance costs.Faced with the challenges posed by the expansion of the railway network and the scarcity of defect samples,particularly the issue of overfitting under few-shot conditions,this study proposes a few-shot deep learning algorithm for the detection of rail surface defects based on multi-scale WideResNet(Wide Residual Network),which employs data augmentation techniques to expand the finite training set,thereby improving the model's generalization capability.By leveraging transfer learning strategies,it applies pre-trained deep learning models to rail defect detection tasks with extracting multi-scale features,reducing reliance on large annotated datasets.The study designs few-shot deep learning model strategies and constructs a metric learning module that enable rapid learning from limited annotated data and effective generalization.Experimental results demonstrate that the algorithm can effectively detect rail surface defects under 10-shot conditions,achieving a model average precision of 83.6%and a recall rate as high as 93.8%.This research provides a new technology for the automated detection of rail surface defects,which is of great significance for enhancing the safety and economy of railway transportation.
作者
李耀
李梅
LI Yao;LI Mei(School of Materials Science and Engineering,East China University of Science and Technology,Shanghai 200237,China;School of Civil Engineering,Suzhou University of Science and Technology,Suzhou 215011,Jiangsu,China)
出处
《华东理工大学学报(自然科学版)》
北大核心
2025年第6期843-849,共7页
Journal of East China University of Science and Technology
基金
江苏省高等学校自然科学研究项目(24KJB560021)。
关键词
铁轨缺陷检测
小样本学习
深度学习
图像处理
机器视觉
rail defect detection
few-shot learning
deep learning
image processing
machine vision
