【作者】 甘胜丰；

【导师】 王典洪；

【作者基本信息】 中国地质大学 ， 地球探测与信息技术， 2013， 博士

【摘要】 钢铁工业飞速发展,带钢生产线的生产速度逐渐提高,同时随着各领域对自身产品质量的要求提高,用户对上游钢铁企业供应的带钢等轧制品也提出了更高的表面质量要求。为此,带钢表面缺陷的在线检测过程对带钢的生产和质量控制显得尤为重要,然而表面缺陷检测的相关技术并未随着图像处理技术的发展而快速发展,其原因是在相关技术转化为生产力的过程中未能更多地结合实际生产的需要做出改进,也不能克服生产环境的变化给这些技术实现带来的障碍。为了进一步提高检测的科学性,为工业的标准化生产提供准确的数据信息,提高生产系统整体的自动化程度,最终突破制约产业转型升级的关键技术；为了提高带钢表面缺陷分类的一致性,保护设备不受缺陷影响,并降低返工成本;为了减少贸易纠纷,维护企业形象与信誉,保持企业核心竞争力,本文对以下内容进行了综述和说明：(1)对国内外带钢表面缺陷检测技术的发展情况进行了综述,并分析了机器视觉带钢表面缺陷检测技术未来的发展趋势。(2)根据不同的生产环节,阐述了适合多种生产环境下带钢表面图像采集的方式和方法,其中包括照明光源、CCD传感器和检测方式等内容。明确了不同工业生产中的多种带钢表面缺陷检测点位置与检测方式。(3)以热轧和冷轧生产中对表面缺陷要求最高的钢种类型为基础,阐明了缺陷形貌、形成机理、位置分布等影响带钢表面缺陷检测与分类结果的主要因素。并根据研究内容提出了三个问题：(1)带钢图像灰度非均匀分布下边界检测的问题。灰度非均匀分布是带钢表面图像常见的状态之一,如果在该状态下边界检测结果的准确率较低,将严重影响带钢边界缺陷的识别率。(2)无固定形态缺陷图像分类问题。多数伪缺陷形貌无固定形态,而且许多缺陷的内在属性也具有无固定形态的特征。如果不考虑缺陷的无固定形态特性,缺陷识别率将无法达到工业要求的结果,影响生产。(3)带钢图像分类方法的实用性问题。分类算法的识别率、噪声型缺陷的过滤能力、分类结果与生产判定标准匹配关系等因素影响带钢图像分类方法的实用性。围绕上述问题,在多种图像检测与分类方法上,做了以下几方面的深入研究工作：(1)研究了带钢表面缺陷检测中典型的图像检测与分类方法以及这些方法适用的范围。(2)研究出基于高斯模型的动态边界检测方法,该方法能够检测带钢背景与表面的灰度差异,动态地确定干扰的灰度范围,与传统带钢表面缺陷检测使用的的固定闽值法和对比度匹配法相比,它能更好的检测边界含重度干扰的边界图像,在其他两种算法准确度仅为4.6%和1.6%时,该算法可以达到62.8%,且实验条件下的检测速度为每千行1.82秒,满足工业要求。(3)在缺陷分割阶段,研究出基于带钢图像灰度标准化的感兴趣区域搜索算法(Regions of Interest, ROI)和基于缺陷距离阈值的缺陷标记与合并方法。新感兴趣区域搜索方法结构简单,且可以通过判别阈值参数调整搜索结果；新的缺陷标记与合并方法可以通过控制4个参数的大小改变缺陷标记和合并区域。实验验证,该方法能对边缘分割后的纹理不连续性缺陷进行重新组合,实现纹理非连续性缺陷地完整分割,以及同一感兴趣区域不同缺陷的分离。在缺陷分类阶段,研究出一种以Isomap算法为基础,适合监督分类的dls-Isomap算法。该算法对roll-swiss数据建立邻域图时能够减少Isomap算法因k值设置不当引起的“短路边”问题,能够解决dbt-Isomap算法无法对多类roll-swiss数据建立完整邻域图的问题,能够用于多类数据的分类工作。实验表明,基于dls-Isomap的分类方法对冷轧带钢表面缺陷的整体识别率可以达到78%,针对热轧带钢的表面缺陷识别率可以达到93%。最重要的是,当缺陷图像中包含较多无固定形态缺陷时,dls-Isomap方法要明显优于其他方法。(4)研究出一种新的图像分类系统,系统由一套图像类型主观评价机制、两组主分类器和一个三级图像分类构架组成。实验验证,噪声型缺陷的过滤错误量占总数量的比例低于1%,多卷带钢的整体识别率达到了90%以上。研究成果已经成功应用于江苏沙钢集团有限公司1450热轧成品机组。经核算,研究应用新增约351.4万元／年的经济效益。更多还原

【Abstract】 There are two points that making the online strip surface defect detection greatly important on steel production and quality control. One point is,with the high speed development of iron and steel industry, the strip production line speed is increasing; another is consumption enterprises on the strip also put forward higher surface quality requirements with their increasing quality requirements in many filed. But steel strip surface defect detection is slow development with the rapid development of image processing technology. The reason of this problem is failure to make more improvements of related technology for production needs and failure to change these technologies for overcome the obstacles form production environment.The overview and research of this paper listed as follows are in order to further improve the scientific detection that including accurate data of Industrial production standardization, to improve the degree of automation, to classification accuracy which related rework cost and maintenance cost, to reduce the trade disputes and keep the enterprise reputation.(1)The developments of domestic and foreign steel surface defect detection technology are reviewed, and analyzed the strip surface defect machine vision inspection technology future trends.(2)Studied for strip surface image acquisition ways with a variety of production environments. The research topics include lighting, CCD sensors and detection methods. And then the Strip surface defect detection position and detection method in many different production environments have been identified.(3)Studied for the main factors affecting the cold/hot rolling strip surface defect detection and classification results, which including morphology, formation mechanism, location distribution.According to the research, three further questions are raised.(1) The question of strip boundary detection under conditions of non-uniform distribution of gray in strip surface image. non-uniform distribution of gray is the common status of strip surface image. Therefore the accuracy of boundary detection results will seriously affect the boundary defect rate of the strip.(2)The question of the classification with no-fixed shape defect images. Most pseudo-defects have no-fixed shape, and many features of defect image also have no-fixed shape. If not consider the non-fixed shape characteristics of defects, the defect rate will not adapt the industrial level.(3)The issues of the strip image classification method have not enough capability in practice.This paper has done the following in-depth research in many image detection and classification methods with these problem:(1)Studied for the typical image detection and classification method in strip surface defect detection and classification.(2)Developed an adaptive boundary detection method which based on Gaussian. This method can detection the gray differences with the surface and the background, can dynamically determine the interference range of gray. Compared to the traditional method of strip surface defect detection, this method can detect stripe boundary that contains the boundary interference. The accuracy of the method up to62.8%form4.6%and1.6%in the most severe case. Detection speed is every thousand lines of1.82seconds(3)In the stage of defect segmentation, developed an Region of Interest algorithm that base on standard strip gray image. This algorithm has the advantages of simple structure, and through the discrimination threshold parameters can be adjusted for search results. Developed an defect image Connected Component marking and merging method by controlling the4parameters can be resized results.In the stage of defect classification, developed double-limited and supervise-connect Isomap dimensionality reduction and classification method (dls-Isomap). Based on the dimensionality reduction technique from Isomap, the connection of neighborhood graph is limited by key parameters k-nn and ε-radius, and inter-class neighborhood points are connected extensionally with the supervision of class labels. According to multi-classes roll-swiss data experiments, all the points can be embedded in lower dimensions with the complete inter-class and intra-class geometric structure, and the "short circuit" in the Isomap can be solved by the dls-Isomap method. In addition, stripe surface defect images data experiments show that the new proposed classification method is suitable for the classification of stripe surface defects including more no-fixed shape defect images, with the recognition rate of78%for cold-roll strip images, and93%for hot-roll strip images with water.(4)Developed a new image classification system, which including a set of image class subjective evaluation mechanism, two sets of the main classifier and a set of three image classification framework. The experiment result show that the filtering error amount of defects in the total number is less than1%. The overall recognition rate of all the experimental steel coil up to90%.The research results have been successfully applied in Jiangsu Shagang Group Co. Ltd. in 1450Hot Rolling Product Unit. The economic benefits of these application about three million yuan per year.更多还原