【作者】 盛党红；

【导师】 黄文良；

【作者基本信息】 南京理工大学 ， 机械电子工程， 2009， 博士

【摘要】 机器视觉是现代制造的一个及其重要的组成部分,涉及人工智能、神经生物学、心理物理学、计算机科学、图像处理、模式识别等多个领域的交叉学科。机器视觉实现制造过程中的运动目标检测和智能控制已成为现代制造领域的研究热点,例如,自动线生产和装配监测、机器人和机械手引导、产品检测和分类、视觉伺服系统、零件图像的自动理解和识别等。因此,现代制造中运动零件图像跟踪和识别研究有着十分重要的意义。本文在分析已有的运动目标跟踪与识别算法的基础上,以机械运动零件图像为对象,将改进遗传算法、神经网络、小波变换、数学形态学的基本定义和基本算法、证据理论有机运用,研究新的机械运动目标跟踪与识别算法,主要研究工作包括如下三个方面:首先,针对成像系统在三维场景转换成二维图像的空间变换过程中存在的非线性几何失真问题,研究了图像非线性失真原理,提出了基于改进遗传算法优化神经网络进而实现图像非线性几何失真的校正算法,实验结果表明了该算法能够增强神经网络的全局搜索能力,提高了收敛速度和稳定性,能够较好地对图像在空间变换过程中存在的非线性几何失真进行校正。其次,探讨基于运动分析和基于图像匹配分析的机械运动目标的跟踪方法,深入分析了帧间差分法、光流法和典型的自由型变形Snake模型,重点解析了帧间差分法因背景变化而造成质心坐标的不稳定性、光流法所存在的背景遮档及孔径问题和Snake的收敛性等问题。提出了基于形态学的机械运动目标跟踪方法,实验结果表明应用数学形态学的基本定义和基本算法的目标跟踪方法为有效地检测出机械运动目标,正确获取运动目标的质心坐标,实现运动目标跟踪提供了可行的方案。最后,研究了机械运动零件图像识别算法。充分利用图像像素之间的空间相关信息,提出了基于Hilbert-小波扫描的图像分割方法,提高了图像分割的效率;应用小波变换方法进行了图像的边缘检测,较好地处理抑制了噪声和边缘定位的矛盾;将被分割的图像和边缘图像划分为子矩阵图像,以获取相对像素系数作为特征向量,将相对像素系数作为神经网络的输入样本,由神经网络实现识别,大大降低了运算的工作量;针对图像传感器在获取零件图像时,由于传感器固有的缺陷、环境因素的影响,有可能难以获取图像较完整的信息,不利于机械零件图像的特征提取和零件的识别,提出了根据证据理论的融合推理规则识别零件的方法。利用LabVIEW软件平台,设计了零件图像识别虚拟仪器,实验结果表明了论文的设计思想和方法达到了预期结果。更多还原

【Abstract】 Machine vision is one of the most important parts in modern manufacture. And it involves multi-region intersection subjects including artificial intelligence, neurobiology, psychophysics, computer science, image processing, pattern recognition and so on. Machine vision achieving moving target inspection and intelligent control has become hot research in modern manufacturing field, such as automation line production and assembling monitoring, robot and manipulator guiding, product testing and classification, vision servo system, automatic understanding and recognition of part image and so on. Therefore, the tracking and recognition research of moving part image is important to modern manufacture.In this paper, on the basis of analyzing existing tracking and recognition algorithms of moving part object, new algorithms of mechanical part moving object have been studied by combining improved Genetic Algorithm, Neural Network, Wavelet transformation and the basic definitions and algorithms of Mathematic Morphology. The main researches include three aspects.Firstly, for solving the problem of nonlinear geometry distortion that three-dimensional scene is converted into two-dimensional image in image-forming system, the nonlinear distortion principal is studied and the correction algorithm based on an improved genetic algorithm optimizing neural network to implement image nonlinear geometry distortion is proposed. The experiment results show the proposed method could enhance the global searching capability of neural network, improve the convergence speed and stability and calibrate image nonlinear geometry distortion in the processing of space transformation.Secondly, tracking methods of mechanical part moving object based on motion analysis and based on image matching analysis are investigated. And inter-frame difference method, optical flow method and typical free deformation model (snake) are analyzed. The problems of centroid coordinate instability resulted from background change in inter-frame difference method, background covering and aperture in optical flow method and the convergence in snake are mainly analyzed. Tracking method of moving object based on Morphology is proposed. The experiment results show the tracking method based on the basic definition and algorithm in Mathematics Morphology provide feasible scheme for effectively detecting moving object, correctly acquiring its centroid and finishing its tracking.Finally, the recognition algorithms of mechanical moving part image are studied. In order to decrease calculation, the image segment method based on wavelet-Hilbert by adequately using the space relative information of pixels is proposed and it improves the efficiency of image segment. In the meantime, the wavelet transform method is also used to detect image edge and solve the contradiction between suppressing noise and edge location. Then, the images segmented and edge images are divided into sub-matrix images in order to get relative pixel coefficients as eigenvectors. And the relative pixel coefficients are the input sets of neural network and the recognition is carried out and it greatly reduces calculation. Because of the inherent drawbacks of sensors and the affects of environment factors, it is difficult to get the complete image information of mechanical part. It makes the feature extraction of part image and part recognition inaccurate. In order to solve this problem, the recognition algorithm based on merging reasoning rules in evidential theory is proposed. In the end, the virtual instrument of part image recognition is designed by using Labview software platform and the experiment results verify the design idea and methods of the dissertation reach expectant results.更多还原