【作者】 张洪涛；

【导师】 叶声华； 段发阶；

【作者基本信息】 天津大学 ， 测试计量技术及仪器， 2008， 博士

【摘要】 钢板表面缺陷在线检测技术已成为钢铁生产企业向现代化制造业高效率、自动化与智能化方向发展的制约因素,同时也是国内外学者研究的热点领域。本文依托科技部科研院所专项基金项目——“钢板表面缺陷计算机视觉在线检测系统研制”,针对系统宽幅面、简单性、模块化、先进性等技术要求,结合国内外表面缺陷检测技术发展方向,确立了以计算机视觉检测技术线阵CCD扫描法为检测原理的钢板表面缺陷检测系统研制方案,并对检测系统关键技术进行讨论。本文主要研究内容与完成工作如下:1.设计了基于线阵CCD扫描、幅面分割的钢板表面缺陷在线视觉检测系统方案,以满足高速、宽幅面、高分辨率的检测要求,同时针对钢板表面缺陷特点,通过优化配置明、暗域成像模型,以检出各类形态缺陷;2.为解决高速运动场景下光源照明问题,设计了狭缝式高频荧光灯光源,其具有亮度高、均匀性好、稳定性高及成本低等特点;3.针对钢板表面缺陷特点及系统指标要求,充分讨论了钢板表面缺陷图像处理算法流程,并针对各处理步骤给出了实现算法。系统可检出钢板主要缺陷,如孔洞、夹杂、划痕、氧化铁皮、斑点等。通过算法优化及采用多线程程序结构,在满足精度指标要求情况下,其处理速度能满足钢板2m/s的运行速度要求;4.在仔细分析钢板表面缺陷特点基础上,本文定义了缺陷特征参数空间,并讨论了决策树、贝叶斯分类器和神经网络分类算法,重点讨论了BP神经网络算法,其缺陷识别率达到90%。由于各种分类算法的局限性,本文也重点讨论了集成分类器,集成分类器通过各种分类器的优势互补,可有效分类各种缺陷;5.提出了整套软件系统框架结构设计方案,即采用多进程架构、基于流水线工作原理和实时采集+准实时处理融合的技术方案。采用优化的多线程程序结构设计采集应用进程,使其数据采集率达到100Mbps,完全满足系统检测要求。为减轻图像准实时处理系统的负荷,实时采集应用进程不仅完成图像数据采集,而且同时进行ROI(Region of Interest)检测;准实时处理应用进程处理ROI图像文件,并提取各个缺陷的特征数据以进行模式分类;6.设计并实现了两套实验样机系统:平动低速模拟系统与高速转动模拟系统,实验中运行的最高速度为1.5m/s。最后通过实验验证了系统方案的可行性。本课题在2007年1月顺利通过国家科技部专家组验收。更多还原

【Abstract】 The technique of on-line surface defects detection for steel plate has become the restriction of the Steel and Iron production enterprises developing to the modernized manufacturing industry with high efficiency, intellectualization and automation, so it’s been a hot field researched by foreign and domestic scholars. Sponsored by scientific research academe specific fund of NSTS, the project is developed to detect surface defects for steel plate based on computer vision. In view of systematic specifications, such as broad width, simplicity, modularization, advancement etc, and combining with the latest development of relevant theory and technology, the scheme of linear CCD scanning method based on computer vision is brought forward to realize on-line non-destructive surface inspection for steel plate, and key technology of system is discussed. The main work included in the dissertation is shown as follows:1. In order to satisfy the requirements of high speed, broad width and high resolution, the linear CCD scanning scheme is put forward to detect the surface defects of steel plate based on width division using on-line computer vision. Taking the characteristics of surface defects into account, optimal configuration of receiving mode containing bright-field and dark-field is analyzed to make defects effectively detected;2. The slit-type high-frequency fluorescent lamp is designed, solving the problem of luminance under high speed movement, and this light source has some excellent characteristics, such as high brightness, good uniformity, high stability and low cost and so on;3. With a view to the characteristics of surface defects for steel plate and the demand of system performance, the algorithm flow of image-processing applied to the surface defect of steel plate is fully discussed, and feasible algorithm is presented at each processing step in the paper, Main surface defects such as hole, inclusion, rolling skin, scratch and roll-mark can be detected. Under the condition of meeting the precision, the application is realized using the optimized algorithm and multi-thread procedure structure, so its processing speed could satisfy the 2m/s running rate of steel plate.4. The dissertation not only defines spatial distribution of surface defects based on its speciality, but also discusses some pattern recognition methods, such as decision tree, Bayes classifier, nervous network means, and it also makes a specific exposition about BPNN that the rate of classifying defects is more than 90%. Because of the limitation of each classifier, compositive classifier is mainly discussed in the paper, making use of each others’ advantages to effectively classify defects.5. Adopted multi-process scheme including real-time gathering system + near-real time processing system, and based on pipe-line theory, software structure is proposed in this paper. The optimized multi-thread procedure structure is introduced to the capturing application to realize the high speed acquiring, and the rate of capturing data reaches to 100Mbps, so it meets the requirement of system. In order to reduce the load of image-processing application, the real-time application not only captures image data, but also simultaneously detects ROI (Region of Interest) data. Near-real time processing system processes ROI file, and extracts special data of defects to distinguish the type of defect.7. Two kinds of prototype, including low speed translational prototype and high speed rotary prototype simulation systems, are developed. The highest speed is 1.5m/s in the experiment. Feasibility and validation is verified through experiments.This project was checked and accepted smoothly by Ministry of Science and Technology at January 2007.更多还原