基于纹理特征增强的重载铁路钢轨缺陷检测算法

Defect Detection Algorithm for Heavy-haul Railway Rails Based on Texture Feature Enhancement

为实现重载铁路轨道典型缺陷的准确、快速、智能检测,基于深度学习算法,提出一种针对重载铁路钢轨图像的特征加强卷积神经网络模型,研制一套基于机器视觉的便携式轨道图像采集系统;整理创建重载铁路钢轨表面多目标图像数据集,并基于此数据集进行训练,实现裂纹、擦伤、块状损伤、接缝4种典型缺陷目标的智能识别;针对数据集中目标尺度分布不平衡的特点,使用聚类算法重新设置锚框的尺寸和数量;对比分析重载铁路钢轨缺陷图像的纹理复杂性与固有特点,引入加权融合池化模块和纹理特征增强模块对自适应训练样本选择(ATSS)算法进行改进。应用所提算法对重载铁路轨道进行检测,4类典型缺陷目标的全类平均正确率达到85.8%。通过与其他9种检测算法的对比,充分验证了所提算法的有效性。

In order to achieve accurate,fast and intelligent detection of typical defects of heavy-haul railway rails,this study proposed a feature-enhanced convolutional neural network model for heavy-haul railway rail images based on deep learning algorithms,and developed a portable rail image acquisition system based on machine vision.A multi-target image dataset of heavy-haul railway rail surfaces was constructed and subsequently used to train a model for the intelligent recognition of four typical defect types:cracks,scrapes,spalling and joints.For the unbalanced distribution of target scales in the dataset,the size and number of anchor frames were reset using the clustering algorithm.By comparing and analyzing the texture complexity and inherent characteristics of heavy-haul railway rail defect images,a weighted fusion pooling module and a texture feature enhancement module were introduced to improve the adaptive training sample selection(ATSS)algorithm.By applying this algorithm to detect heavy-haul railway rails,the average detection accuracy of all four typical defect targets of heavy-haul railway rail reached 85.8%.Through comparative with nine other detection algorithm,the effectiveness of the algorithm model was verified.