【作者】 赵大兴；

【导师】 颜云辉；

【作者基本信息】 东北大学 ， 机械设计及理论， 2009， 博士

【摘要】 随着生产和工艺的进步,人们对产品的质量要求越来越高,基于机器视觉的在线检测系统成为一种重要的质量控制手段。本文对应用于宽幅面、高精度的基于机器视觉的坯布表面质量在线检测理论与算法进行了研究,研制并开发了一套坯布表面质量检测系统。首先,根据坯布表面质量检测中的宽幅面、高精度的特点,设计了一种主／从机分布式机器视觉在线检测系统结构,采用多个图像处理子系统协同完成检测任务,保证了系统快速处理海量图像的可靠性,并使用千兆以太网来完成图像数据与控制指令的网络传输,保证了图像采集、处理、传输和存储的实时性。文章针对坯布纹理特点,设计了大功率LED条形阵列的正向单侧可变角度照明,光照强、照度均匀,避免了固定角度照明适应性差,影响疵点信息提取的缺点。为滤除坯布图像在采集过程因光照、拍射角度、镜头污染等加入的噪声,分析了根据坯布纹理特征选择滤波器的原则,对频谱特征提取方法、模糊结构元纹理定义以及坯布纹理的分析方法等进行了探讨,研究了基于扇区能量统计的谱特征提取方法,较好地解决了坯布纹理的处理问题。同时,针对滤波执行效率低、复杂一维信号特征提取困难等细节问题,提出了滤波类算子的优化方法,明显提高了算子的执行效率。针对坯布纹理图像的特点,研究了一种适合在非规则纹理图像中寻找目标区域的高精度自适应阈值分割算法。同时,根据疵点特征类型及其区域识别的需要,提出了用矩形窗沿直线扫描,通过计算窗内斑点的总面积来确定是否保留直线上该点的窗线扫描方法,解决了预处理后图像中疵点信息断裂的难题,在此基础上通过两个实例验证了其优越性。然后,根据坯布疵点的常见类型,提取了方差-密度、对比度-密度、熵-密度、频域统计等作为对疵点特征的描述参数,通过求各类疵点特征的并集作为对其所有特征的统一描述,设计了基于特征参数的坯布疵点分类方法,研究了基于改进型BP神经网络的坯布疵点分类器,实现坯布的最终质量评价。最后,在上述算法和理论的基础上,设计开发了坯布表面质量检测系统。实验分析表明,本系统的检测速度最高达200 m/min,疵点的最高检测精度为0.5mm,能够实现对常见8类疵点的实时分类,准确率为90%以上。更多还原

【Abstract】 With the advancement of manufacture and techniques, the product quality is more and more important, and the online quality detection system based on machine vision has been an important quality control method. The paper studies the online quality detection theory and algorithm of fabric surface, which is applied in broad face and high-precision,and based on machine vision. Besides, the paper develops a set of fabric surface quality detection system.At first, due to the characteristics of broad face and high-precision in fabric surface quality detection, this paper designs a kind of structure of master-slave distributed on-line machine vision inspection system which uses multi-image processing subsystem to cooperate detection task, ensures the reliability of dealing with mass image rapidly. In addition, this structure uses Gigabit Ethernet to complete the network transmission of image data and controlling command, which ensures the real-time image acquisition, processing, storage and transmission. Aimed at fabric texture characteristics, the paper designs positive side angle variable lighting of high-power LED bar array, which has strong light and symmetrical illumination, far away from fixed angle lighting’s disadvantages of poor adaptability and defect information extraction.In order to filter noises produced in the fabric image acquisition process by sunlight, shooting angle, camera lens pollution, the thesis analyzes principles of choosing filter according to fabric texture characteristics, discusses the spectrum feature extraction method, fuzzy structuring element texture definition, and the analyzing method of fabric texture, studies spectral feature extraction methods based on energy Sector statistics, and then solves problems of dealing with fabric texture. At the same time, for some detail problems of low filtering performance efficiency and difficult feature extraction process of complex one-dimensional signals, the paper designs the optimizing method of filter type operator, and improves operator’s performance efficiency obviously.For texture characteristics of fabric image, the paper designs the high-precision self-adaptive threshold segmentation which is suitable for finding the target area in irregular texture image. At the same time, based on the need of the recognition of the defects characteristics and its region, the paper designs the window line scanning method with rectangular windowsill in a straight line scan and through the calculation of the total area of the spot in the window to determine whether to retain the point of a straight line. The method solves the problem of the loss of the detects information after pre-processing, and on the basis of it, the superiority is demonstrated by two examples.Then, according to the fabric of the common types of defects, the thesis takes the values of Variance- Density, Contrast- Density, Entropy-Density, Frequency-domain as the basic description parameters for characteristics of defects, unites all kinds of defect feature as unified description, designs fabric defect classified method based on characteristic parameter, researches fabric defect classifier based on mended BP neural network, and realizes fabric’s final quality evaluation.Finally, on the basis of methods and theory above-mentioned, the fabric surface quality detection system is designed and developed. The experimental analysis shows that, the highest detect speed of the system is 200 m/min, the highest detection precision is 0.5 mm. The system can realize real-time classification of common 8 types of defects with 90% accuracy rate.更多还原