【作者】 王鹏；

【导师】 刘献礼；

【作者基本信息】 哈尔滨理工大学 ， 机械制造及其自动化， 2008， 博士

【摘要】 轴承是机械基础部件,钢球作为轴承的滚动体,其表面质量直接影响轴承精度、动态性能和使用寿命,因此,对钢球表面质量的检测技术的研究是具有非常重要意义的课题。本文应用运动视觉技术对钢球表面缺陷进行图像序列获取、分析、跟踪与识别研究,并进行了运动视觉检测仪器的开发,其主要研究内容如下。首先,设计了钢球表面缺陷视觉检测实验系统的整体方案,开发了主要的检测机构模块及实验系统控制单元及电路板,从而搭建了运动视觉检测的实验平台。针对钢球视觉检测系统存在的畸变问题,设计了基于多项校正模型的摄像机标定方法。首先综合分析了钢球序列图像畸变的特征,得出钢球机器视觉检测系统共存在两种畸变:摄像机安装畸变和几何畸变。逐一针对畸变进行研究并建立校正矩阵。然后计算坐标系统的变换和校正后摄像机模型的内外参数,最终推导出系统标定的数学模型。其次,研究了钢球图像序列的去噪和复原问题。在分析现场噪声的函数特性和钢球运动模糊的规律基础上,提出了钢球运动图像退化模型,并设计了基于参数估计的维纳滤波方法进行图像恢复。实验效果表明本文复原算法具有明显的清晰度优势。还针对钢球反光与法线光晕现象,分析了其原因并探讨了两种可能的解决途径。设计了组合式照明方案,实现了钢球反光与法线光晕的弱化。再次,针对场景中多个运动钢球序列图像追踪的问题,设计了基于卡尔曼运动估计的免疫自适应模板追踪算法。首先,利用基于改进的免疫算法搜索出最优相关匹配点,检测出钢球。然后,采用Kalman算法得到每个钢球运动参数的估计,获取运动钢球的位置信息。并对匹配模板编号。最后,以基于加权的自适应模板更新算法解决跟踪点漂移问题。还提出基于蚁群算法的动态轮廓模型(Snake)实现钢球轮廓的自动提取策略。第一步,设计了基于圆弧邻域扩展连接边缘的改进型Canny算法来获得钢球Snake模型的初始化轮廓。第二步,设计了基于时间模型的蚁群算法获取Greedy演化最优解,获取了最佳轮廓。最后,对钢球表面缺陷的有无判断和缺陷分类进行了研究。首先分析了钢球表面图像纹理特征,在研究了能量特征、熵值特征、对比度特征、局部均匀特征、相关性参数的基础上,设计了综合熵参数评判钢球表面缺陷的有无。然后还设计了快速OTSU算法来获得钢球表面的缺陷区域。提出了以形状因子F结合3个不变矩,构成4维特征向量来标示钢球表面曲线区域的特征。最后设计了基于验证更新策略的AdaBoost分类器算法,完成对钢球表面缺陷类型的分类。更多还原

【Abstract】 Bearing is basic part in mechanical equipment, which is used widely. To a large extent, bearing’s precision, motility and service life depend on the quality of steel ball in bearing. Research is very important task for detection surface defect of steel ball. In order to inspect and classify effectively surface defects of steel ball, an automatic measurement system was researched with dynamic vision technology. The main content of this thesis includes the following parts:The whole design scheme and the main module of detection mechanism were described in detail. The thesis also integrally analyzed the control circuit design of automatic detection of steel ball’s surface defect. The motion vision experimental platform was developed for steel ball surface defect detection.Aiming at the distortion of steel ball vision experimental system, the improved traditional camera calibration method was designed based on multi distortion model. First of all, the characteristic of steel ball image sequence distortion was synthetically analyzed, and then the conclusion was got that there were two distortions: the camera installation distortion and the geometric distortion. Then the correctional matrix was established according to various kinds of distortions. The thesis calculated the transformation of coordinate system and the inner and outer parameters of the camera with correction. Finally the mathematic model of system calibration was derived.The problems of noise-removing and restoration of steel ball image sequence were studied.On analyzing the function characteristic of noise-fielded and rule of steel ball motion obscuration, the thesis put forward to the steel ball degradation model and designed the wiener filtering method based on parameter estimation to restore image. The experimental results indicated that restoration algorithm has obvious advantage for articulation in this paper. The thesis also analyzed their optical causes and investigated two possible solutions aiming at the reflection of steel ball and its normal line halo phenomenon .Combined lighting scheme were designed to weaken the reflection of steel ball and the normal line halo.Aiming at tracing image sequence of several motion steel balls in scene, the thesis designed the immunity self-adaptive tracing algorithm based on Kalman. First of all, the initial original template was determined by steel ball gray centroid. Then using Kalman arithmetic, the thesis obtained the estimation of every steel ball motion parameter and the position information. The matching template was numbered. Finally, the problem of tracing point drift was solved based on the weighted self-adaptive template updating arithmetic. The thesis put forward to the dynamic contour model based on the ant colony algorithm to realize the strategy of automatic extraction. On the first step, based on the expanded joint edge of the circular arc neighborhood, we designed the improved Canny arithmetic to obtain the initial profile of the steel ball Snake model.Second, we designed the ant colony algorithm which was based on time model to obtain the best solution of Greedy evolvement and gain the best profile.The thesis studied the judgment whether the steel ball has defect or not and the defect classification.First of all, the surface images texture feature of steel ball was analyzed. Based on the research of energy feature, entropy feature, contrast feature, partially-homogeneous feature and relativity parameter, the synthesis entropy parameter was designed to judge. Then the fast algorithm for Otsu was designed to get the defect area of steel ball surface. The thesis also proposed that by the shape factor F combined with three moment invariants the four-dimensional eigenvector was constructed to sign the feature of steel ball surface curve area. Finally, the thesis put forward the AdaBoost classifier algorithm based on the verification update strategy to classify the defect type of steel ball surface.更多还原