基于图像特征的钢轨表面瑕疵识别方法

发布时间：2018-04-02 10:20

本文选题：图像特征　切入点：钢轨缺陷检测　出处：《西南科技大学》2017年硕士论文

【摘要】：针对现有钢轨表面缺陷检测方法存在适应性差、可靠性不强、自动化程度不高等问题,设计了一种基于图像特征的钢轨表面缺陷检测系统,使用数字图像处理技术与机器学习方法对钢轨表面几种典型缺陷进行判断识别。本文首先介绍了无损检测技术的发展现状,并分析了钢轨表面几种典型缺陷类型以及产生的原因,设计了一种基于图像特征的钢轨表面缺陷检测系统,该系统主要包括图像预处理、特征描述、分类器设计等3个方面。在预处理阶段,首先通过改进投影法提取出钢轨所在区域;其次,通过对噪声类型进行分析,选择使用自适应中值滤波算法对钢轨图像进行滤波操作;针对钢轨表面图像灰度分布均匀的特点,提出一种分块自适应模糊增强算法,根据子块熵值判断,对缺陷潜在子块进行模糊增强,并通过OSTU阈值分割方法实现图像分割;使用空频域相结合的方法,分别提取缺陷图像的灰度、几何、不变矩,以及小波变换后各区域的均值、方差作为钢轨图像的特征;最后,通过设计训练生成BP神经网络模型,来达到钢轨图像表面缺陷检测的目的。通过实验结果分析,该系统可以实现裂纹、划伤、轧疤、凹坑等4种典型缺陷的识别与分类,总体达到漏检率8%,准确率88.5%的指标,实现表明该系统对实际应用具有一定参考价值。
[Abstract]:Aiming at the problems of poor adaptability, low reliability and low automation in existing rail surface defect detection methods, a rail surface defect detection system based on image features is designed. Several typical defects on rail surface are judged and identified by digital image processing technology and machine learning method. This paper first introduces the development status of nondestructive testing technology. Several typical defect types on rail surface and their causes are analyzed. A rail surface defect detection system based on image features is designed. The system mainly includes image preprocessing and feature description. In the preprocessing stage, the rail region is extracted by the improved projection method, and the noise type is analyzed, and the adaptive median filter algorithm is used to filter the rail image. According to the characteristic of uniform gray distribution of rail image, a block adaptive fuzzy enhancement algorithm is proposed. According to the entropy value of sub-block, the defect potential sub-block is enhanced by fuzzy enhancement, and the image segmentation is realized by OSTU threshold segmentation method. The space-frequency domain method is used to extract the gray level, geometry, invariant moment of defect image, and the mean value and variance of each region after wavelet transform as the features of rail image. Finally, BP neural network model is generated by design training. Through the analysis of the experimental results, the system can recognize and classify four typical defects, such as crack, scratch, rolling scar and pit, and reach the target of 8% leakage rate and 88.5% accuracy. The implementation shows that the system has certain reference value for practical application.
【学位授予单位】：西南科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U216.3;TP391.41

【相似文献】