【作者】 朱铮涛；

【导师】 黎绍发；

【作者基本信息】 华南理工大学 ， 计算机应用技术， 2004， 博士

【摘要】 基于计算机视觉图像的精密测量在国外已经得到了深入研究和广泛的应用,在国内也受到了越来越多的重视。随着计算机硬件性价比的不断提高,相关技术的不断发展,计算机视觉图像精密测量以其柔性、快速、非接触、精确、自动化程度高等特点将很快得到全球化应用。本文以图像为基础,深入研究基于计算机视觉图像实现精密测量的系统方案、软硬件设计、特别是以软件为主要手段解决视觉测量中精度、速度和稳定性三大问题的关键技术。文章分析了计算机视觉图像精密测量的基本原理、硬件选择原则、软件功能模块和检测流程,推导了相对标定法的测量精度,从而证明计算机视觉图像测量实现高精度测量的可能性。针对视觉测量的特点,研究彩色图像灰度化方法,提出了彩色图像灰度化效果好坏的评判标准,给出了阶跃边缘、屋顶边缘保持边缘结构特征的评价函数。根据视觉测量图像质量较好及以高斯噪声为主的特点,提出采用简单的高斯滤波或者SUSAN 去噪算法对图像在卡尺范围内进行滤波处理的方法。提出了基于知识的边缘检测方法,利用卡尺技术,在知识库的引导下,利用SUSAN 算子中USAN 的取值规律和边缘判别准则提取像素级边缘,提高边缘检测中像素级边缘的定位精度、抗噪声能力和运算速度。结合SUSAN 和Canny 算子各自的优点,提出了基于Canny 和SUSAN 算子的边缘检测方法,具有较好的边缘检测效果。以理想一维阶跃边缘模型为基础,对基于矩的亚像素定位算法的定位精度进行了深入的研究。指出空间矩的边缘定位精度敏感于采样间隔、计算点数、边缘对称性。首次提出在使用灰度矩计算亚像素边缘位置的时候,正确区分边缘类型,在两种极限状态下(两个像素之间和1 个像素中间),选取对称点进行灰度矩亚像素定位计算,理论上可以实现无偏的定位估计。提出了基于前后向差分和曲线拟合的亚像素定位算法,理论上该算法与灰度矩算法一样,在两种极限状态下可以实现无偏的边缘定位估计,其它情况下的定位精度要好于灰度矩。研制了一种用于镜头畸变校正的平面圆形网格模板,推导了利用该模板计算镜头畸变系数的公式,提出了计算机视觉图像精密测量标定的两步法:实验室计算镜头畸变系数,现场进行像素值和测试数据计算。由于只进行几何畸变校正,每一步均为线性矩阵运算,解决了通用计算机视觉摄像机标定中速度、精度之间的矛盾。分析和对比了各种自动聚焦判决函数的特点,在均方差自动聚焦函数的基础更多还原

【Abstract】 Computer vision image precision measurement has been researched further and applied widely abroad as well as popular at home. With the ratio of computer hardware and performance increasing, the relative technology developing, the character of flexibility, fast speed, non-contact, high precision and automatization, the computer vision image precision measurement application will become global. Based on image, the system scheme of the computer vision image precision measurement, the design of software and hardware, especially the key technologies of solving precision, speed, stability using software have been researched further. This paper analyzed the basic theory of computer vision image precision measurement, hardware selecting, software function module and inspection flow, deduced the measurement precision of relatively calibration. The possibility of realizing the high precision measurement has been proved. According the characteristic of computer vision measurement, the author researched the methods of color image converting to gray and presented judging rules of gray transform arithmetic. The judging functions of keeping the edge structure characteristic have been presented aiming to step edge and roof edge. The simple GAUSS filter or SUSAN filter will be used to removal the noise in the calipers scope according the characteristic of image quality and GAUSS noise. This paper presented the edge detection method based on knowledge,utilized the calipers technology and USAN rule of SUSAN arithmetic to locate edge to pixel precision with the knowledge. It will improve the edge locating precision, resist noise and improve the calculating speed. Combining the merits of SUSAN and Canny, this paper presented the method of edge detection based on SUSAN and Canny. The method has the better results of edge detection. The sub-pixel edge location technology using moment were studied further in this paper based on the 1-D ideal step edge. The paper presented that the locating precision of spatial moment relates the sampling interval, calculating counter and edge symmetry. The gray level moment edge locating can acquire no error estimate when we do not consider the noise and selecting the symmetry edge in two limit states (inter-pixel edge and intra-pixel edge). The author presented the method of sub-pixel edge locating based on back and forth difference and curve fitting. This method has the same locating precision as the gray level moment in the two limit states and has the better result at the other condition. This paper developed a plane rotundity gridding template to proofreading the lens distortion. Using this template we can calculate the lens distortion coefficients. The author presented the two steps calibration for the computer vision image precision measurement, calculating the coefficients of distortion in the lab, pixel value and test data in the scene. As there are only linear matrix calculation and geometry distortion proofreading, this method solved the problems between the speed and precision in computer vision calibration. This paper analyzed the auto-focus judging function and presented the mean-square judged-function weighted in condition on the basis of mean-square difference judge-function. Selecting the appropriate scope in the computer vision image precision measurement, it will be reach the good focus effect. The technology problems of realizing the precision measurement using computer vision have been solved basically when we use the image in focus, the high precision edge location and calibration technology, distortion proofreading, callipers and knowledge. It is possible that we will use the computer vision widely in industry. This has important significance to promote the application of vision test technology.更多还原