【作者】 张浩鹏；

【导师】 王宗义；

【作者基本信息】 哈尔滨工程大学 ， 检测技术与自动化装置， 2013， 博士

【摘要】 视觉检测具有速度快、非接触、较高精度、操作简单以及自动化程度高等优点，因此视觉检测已经在许多领域内发挥巨大作用。本文对视觉检测系统的若干关键问题进行了研究，主要研究内容可以概括为以下几个方面：1、为了克服在摄像机标定过程中需要使用者给出标定模板的附加信息，或全自动标定点识别算法在遮挡、不均匀照明、大视角和摄像机镜头畸变情况下不能检测出标定点的缺点，提出一种改进的基于基准点标记的棋盘格模板以及相应的全自动识别算法。新的摄像机标定模板以基准点标记代替传统棋盘格的黑白方块，从而使全自动识别算法识别出标记的位置。利用模板中标记按照标记ID从小到大的顺序排列的先验知识，估计丢失的标定点位置。为了提高丢失标定点在图像中初始位置的估计，算法估计径向畸变参数，从而克服了畸变对识别的影响。为了提高标定点的定位精度，利用高精度的鞍点检测器。为了检测鞍点的有效性，算法提出2种滤波准则，最终得到有效的标定点。识别算法是有效的且不需要任何参数。2、传统摄像机标定同时计算针孔模型和透镜畸变模型，从而两种模型耦合在一起，所得到的标定结果仅对训练数据是有效的，而算法对于新数据的标定误差增加。为了克服两种模型的耦合，一种分离地标定两种模型的方法被提出。算法利用模板的射影不变量约束，即交比和直线的射影为直线，求解带有畸变的标定点的校正的坐标，然后线性地求解两种模型。该算法稳定且精度高。3、图像传感器通常受到脉冲噪声的干扰。为了从获取的数字图像中去除脉冲噪声，且同时保留图像的边缘特征，提出基于灰度一致性的脉冲噪声滤波器。滤波器由脉冲噪声检测器和自适应开关中值滤波器组成。由于脉冲噪声是固定值脉冲噪声和随机值脉冲噪声的混合，所以检测器首先利用噪声图像直方图去除固定值脉冲噪声，然后利用方向模板检测待测点是否同任一方向的像素满足灰度一致性准则，从而找到任何可能的随机值脉冲噪声污染的像素。随后，自适应开关中值滤波器利用二进制检测图判断是否进行滤波和确定滤波窗口尺寸，采用窗口内无噪声像素的中值替换检测出的噪声像素。为了获得最优的滤波结果，提出自适应地确定迭代数目的滤波方案。提出的滤波器能够滤除脉冲噪声和具有较快的处理时间。4、分析原始Canny边缘检测算法的不足，提出改进的Canny边缘检测算法。该算法首先利用小尺度的高斯函数平滑图像，提高边缘检测算法的定位精度，然后利用Otsu双阈值算法自适应地计算梯度幅值的双阈值，利用Otsu单阈值算法自适应地计算灰度的阈值，灰度的阈值对梯度幅值的双阈值进行更强的限制，更加准确地寻找边缘，排除无意义的边缘。改进的Canny边缘检测算法避免人为设定参数，适合工业检测的需要，具有很好的抗造性能、边缘定位精度和实时性。5、分析伪边缘的性质，提出判断准则，排除明显错误的伪边缘，提高边缘匹配的速度和精度。利用基于弦到点距离累积技术的边缘特征点检测算法检测剩余边缘的特征点。利用各种约束限制候选匹配特征点的搜索范围，然后利用边缘描述子找到正确的匹配点。边缘描述子由灰度均值，灰度标准差和边缘方向直方图组成。边缘描述子具有优良的匹配性能，并且对遮挡、平移、旋转和线性光照具有不变性。匹配特征点所在的边缘是匹配的，利用边缘约束范围，利用外极线约束实现边缘点的快速匹配。6、分析精度造船对船体分段视觉检测系统的要求，即快速和高精度地获取船体分段边缘的三维数据，从而提出船体分段视觉检测系统的硬件设计方案。硬件由机械伺服系统、视觉传感器、图像采集系统和计算机组成。本文给出硬件的选型，并着重分析视觉传感器的结构设计。同时给出船体分段扫描的实现方法，包括基于移动坐标的测量方法，基于位置触发的船体分段扫描方式和摄像机和图像采集卡的触发方式。高性能的硬件保证船体分段视觉检测系统的测量精度和测量速度，高精度的检测结果保证船体分段无余量对接和补偿量系统的建立。更多还原

【Abstract】 With the advantages of fast measuring speed, non-contact, high accuracy, easy operationand automatic measuring, etc, the stereo vision inspection technique has played a significantrole in many areas. We discussed some key problems about the stereo vision inspectionpractically and theoretically. The major content is as follows:1、In order to overcome camera calibration needs user to give additional information ofcalibration pattern, or fully-automatic identification algorithm of calibration points can’tdetect calibration points in the presence of significant occlusions, uneven illumination,observations under extremely acute angles and lens distortion, an improved checker patternbased on fiducial markers is designed, and its correspondingly fully-automatic identificationof calibration points is proposed. New camera calibration pattern replace black and whitesquares of traditional checker pattern with fiducial markers, so fully-automatic identificationalgorithm can locate markers. Using the priori knowledge that markers are arrangedsequentially into calibration pattern in accordance with markers’ ID from small to large, themissed calibration points can be located. In order to improve the estimate of initial position ofmissed points in image, radial distortion is estimated, so identification algorithm overcomethe impact of lens distortion. In order to improve location accuracy of calibration points,subpixel-accurate saddle points’ finder is used. In order to validate the potential points, twofiltering criteria is performed, finally valid calibration points is obtained. The detectionalgorithm is efficient and free of parameters.2、A lot of traditional camera calibrations compute the pin-hole and lens distortionmodels at the same time, so two models are coupled, computed calibration results are validonly for training data, and their calibration errors increase for new data. In order to overcomethe coupling of two models, a calibration which calibrate two models separately is proposed.The algorithm uses geometric invariants of calibration template, which are cross-rationinvariability and the projection of straight line in the space to the camera is a line, to computeundistorted coordinates of distorted calibration point, and then compute two models linearly.The proposed algorithm is stable and has high calibration accuracy.3、Vision measurement system is usually subjected to the interference of impulse noise.In order to remove impulse noise from the obtained digital images and also preserve image edge features, impulse noise filter based on gray consistency is proposed. The proposed filteris composed of an impulse noise detector and an adaptive switching median filter. Becauseimpulse noise is some mixture between the fixed-valued impulse noise and the random-valuedimpulse noise, the detector first utilizes the noise image histogram to identify the fixed-valuedimpulse noise, then utilizes direction templates to detect whether detected point meet thecriterion of gray consistency with pixels in any direction, and classifies any possiblerandom-valued impulsive noise pixels. Subsequently, adaptive switching median filter usesbinary decision map to determine whether the filter is implemented and determine thedimensions of filtering window, and replaces the detected noise pixels with the median pixelof all noise-free pixels in filtering window. In order to obtain the optimum performance, thefiltering scheme that adaptively determines the number of iterations is proposed. Theproposed filter is capable of filtering impulse noise and has the relatively fast processing time.4、Analysis the deficiencies of original Canny edge detection algorithm, the improvedCanny edge detection algorithm is proposed. First, the algorithm use small-scall Gaussianfunction to smooth image, improve location accuracy of edge detection algorithm. Second, ituse Otsu dual-threshold algorithm to calculate the dual-threshold of gradient amplitudeadaptively, use Otsu single threshold algorithm to calculate the threshold of grayscale, whichlimits the high threshold of gradient magnitude, more accurately finds the edge, excludes themeaningless edge. The improved canny edge detection algorithm avoids set parametersartificially, meet the needs of industrial inspection, has good anti-making performance, edgelocation accuracy and real-time.5、Analysis the nature of error edge, propose the criterions to exclude error edges,improve the speed and accuracy of edge matching. Use the corner detector based on thechord-to-point distance accumulation technique to detect the festure points of remainingedges. Use a few constraints to limit the search of candidate matching feature points, then useedge descriptor to find right matching point. Edge descriptor is composed of grayscale mean,grayscale standard deviation and edge direction histogram. Edge descriptor has excellentmatching performance, and is invariant to occlusion, translation, rotation and linearillumination. The edges including matching feature points are matched, use edge to constrainthe range, and use epipolar constraint to achieve fast matching of edge points.6、Analylysis the requirements of accuracy shipbuilding for hull section visual inspectionsystem, that is to obtain the three-dimensional data of the edge of hull section, so thehardware design of the hull section visual inspection system is proposed. Hardware iscomposed of mechanical servo system, vision sensor, image acquisition system and computer. This paper gives the selection of hardware, and focuses on analysis of the structural design ofvision sensor. At the same time, it gives the implementation method of hull section scanning,including measuring method based on moving coordinate, hull section scanning mode basedon location trigger, trigger mode of camera and image acquisition card. High-performancehardware guarantees the measurement accuracy and speed of the hull section visual inspectionsystem. High-precision detection results guarantee the establishment of marginless dockingand compensation system about hull section.更多还原