【作者】 张静；

【导师】 叶玉堂；

【作者基本信息】 电子科技大学 ， 光学工程， 2013， 博士

【摘要】 表观缺陷检测技术是印刷电路板(Printed Circuit Board，PCB)行业提高产品生产力，改进生产工业化水平的关键技术之一。随着计算机科学、图像处理、模式识别等诸多领域的发展，基于机器视觉的自动光学缺陷检测技术取代了传统的人工目检技术，对采集到的光电图像提取有用信息，进行处理并加以理解，最终完成实际检测，该技术成为PCB表观缺陷检测今后发展的方向。在上世纪末，自动光学检测(Automatic Optical Inspection，AOI)技术在国际上已经得到一定程度的应用和推广，我国对PCB表观自动光学检测技术起步较晚，一定程度上影响了国内PCB产品的质量评价和市场竞争力。由于受到表观缺陷检测质量、检测速度等方面因素制约，整体系统仍处在研究和发展阶段。如何降低系统复杂度，增强系统稳定性，降低系统成本，优化缺陷检测和分类方法，提高检出率和分类正确率，成为自动光学检测的研究重点。本文针对PCB表观缺陷AOI系统的设计、缺陷的识别分类所涉及到的相关理论及关键技术进行了深入的研究，主要成果及相关研究内容如下所述。(1)对PCB表观缺陷自动光学检测系统技术指标和性能进行分析，将系统按照功能划分为不同模块，提出针对PCB表观检测的照明、硬件和软件设计方案，研究利用空间评价函数调整照明最佳配置，采用明暗域结合的照明方式，配合高速线阵CCD采集图像。对硬件设备进行功能参数匹配以保证运动状态下图像质量清晰稳定。按照检测需求对软件进行模块化的系统设计，采用工位划分的流水处理模式，加入并行编程的设计架构，最终实现智能化的自动光学检测系统。(2)在预处理单元，由于受制作工艺和照明设备的影响产生色彩偏差。分析色彩空间模型，研究CIE-Lab色彩模型中的亮度通道，利用亮度累积直方图计算映射函数去均衡板内的色差；利用照明模型得到亮度变换函数对板间色差进行校正。针对工业检测图像噪声来源的多样性，研究结合形态学的自适应各向异性扩散方程滤波去噪的方法，提高了检测的精度和系统性能。(3)在图像配准和标准板的制作单元，考虑到PCB定位孔图像出现的部分缺失及图像配准对精度的要求，提出了一种基于随机Hough变换和空间数据坐标变换相结合的配准方法，该方法有效地提高了配准效率及精度，减少了计算时间。本系统首次将描述电气物理性质的Gerber文档引入建标过程，利用正则表达式自上而下的分析方法解析不同型号电路板对应的Gerber文档，通过形态学和神经网络的算法对解析后的图像进行修正，从而得到精准的PCB标准板，为后续缺陷检测和分类奠定良好的基础。(4)在PCB表观缺陷的特征提取单元，由于受到电路板不同材质及缺陷形成机理的影响，不同物理层上的缺陷区域存在过渡区，提出将分形维数和过渡区理论相结合的局部动态阈值分割方法。该方法结合缺陷的过渡区域信息，利用分形维数对分层后的图像划分不同的子图区域，弥补局部阈值分割方法中子图像大小影响最终分割效果的问题，最终采用动态阈值进行图像分割，提高了缺陷提取的完整性和准确性。(5)在对提取到的表观缺陷进行分类单元，研究了局部二元模式(LBP)和图像颜色特征相结合的纹理算子LBPC，通过卡方公式计算缺陷样本训练集和测试集的特征分类距离，从而完成对缺陷种类的自动分类识别。实验结果表明表观缺陷分类准确率得到明显提升，通过与传统自适应神经网络分类算法进行对比，分类正确率提高了12%，达到95.5%，满足了工业生产的需要。(6)在对整个算法系统进行加速阶段，研究利用图形处理单元(GraphicProcessing Unit，GPU)对设备的实时性效果进行改进，深入分析CUDA的设计模式带来的并行处理优势，研究了利用该技术对PCB自动光学检测系统复杂算法的改进。通过实验结果分析表明系统实现了并行处理对表观缺陷图像的预处理、缺陷提取和自动分类算法的加速，大大缩减了整个系统的运行时间，对于数据量较大的图片，运算速度能提高近30倍左右。本文通过对PCB表观缺陷自动光学检测技术理论和关键技术研究，提出了表观缺陷检测系统的设计方案、图像处理算法及分类识别方法，并且利用计算机图形图像处理单元降低算法系统的复杂度，提高图像处理的执行效率，完成25cm×22cm整板PCB缺陷检测平均时间仅需要3s，大大改善了自动表观检测系统的实时性，在实际的工程项目中得到验证，目前已成功应用在工业检测领域。更多还原

【Abstract】 The apparent defect detection technology is one of the key technologies forimproving product productivity and increasing production level of industrialization inprinted circuit board (PCB) industry. With the development of computer science, imageprocessing, pattern recognition and etc., automated optical defect inspection technologybased on machine vision replaces the traditional manual visual inspection technology.The useful information is extracted from optical imagest to process for understanding.The actual detection results are ultimately finished. The technology is the maindevelopment direction of PCB apparent defect detection in the future.In the later of last century, automatic optical inspection technology has been acertain degree of application and promotion in the world. But in our country, there isstill very large gap from others, which affectes the quality of PCB products and marketcompetitiveness to a certain extent. Due to constrain of apparent defect detection quality,detection speed and etc., overall systems are still in the research and development stage.It is automatic optical inspection (AOI) research focuses that how to reduce systemcomplexity, enhance system stability, reduce system costs, optimize defect detection andclassification methods and improve the detection rate and classification accuracy. In thispaper, PCB apparent defects AOI system design and defect identification classificationand other key technologies are researched. The main achievements and relative contentsare listed below:(1) The technical indicators and performance of apparent defects on PCB AOIsystem were analysised, according to the system function, the system into differentmodules are divided, the lighting, hardware and software design scheme for PCBapparent detection are proposed. Lighting optimal configuration was adjusted by usingspace evaluation function, an illumination of light and dark field combination andacquired image with high-speed linear CCD is adopted. To ensure the image clarity andstability in motion, hardware devices are matched with each function parameters. Inaccordance with the inspection requirements, a modular software system with waterprocessing mode is designed, and ultimately satisfies the need of intelligent AOIsystem.(2) In the preprocessing unit, owing to the manufacturing process and lighting effects, the system is prone to bring color deviation. Therefore, color space models areanalyzed, using the luminance channel of CIELAB color model and utilized mappingfunction by accumulated luminance histogram to balance the plate colors. The lightingmodel is used to get the brightness transformation function and then corrects the platecolors. For the noise sources diversity of industrial inspection image, with morphologyadaptive anisotropic diffusion equation, partial differential equations are studied to filternoise. The experimental results show that the method is good at reducing noise, whileretaining the image edge information. This approach protects the credibility of thepost-processing results and improves the detection accuracy and system performance.(3) In the image registration and establishing standard board unit, considering thevarious forms of PCB positioning hole and image registration accuracy requirements,we propose a novel registration approach based on randomized Hough transform andspatial data transformation theory. The method has better performance in finding thetarget, improving detection accuracy, reducing memory space and the computationaltime. The Gerber file into building Standards process is introduced; regular expressiontop-down is used to analyze the Gerber files corresponding to different circuit board. Bymorphology and neural network algorithm to rectify the parsed image, accurate PCBstandard board to lay a good foundation for subsequent defect detection andclassification are obtained.(4) In feature extraction of PCB apparent defects unit, because of the circuit boardsof different materials and various defects, there has a transition zone in differentphysical layer defect area. Thus the local dynamic threshold method combining fractaldimension with the transition zone theory is proposed. The method combined defecttransition zone information uses fractal dimension to divide layered images intodifferent sub-plot areas. It can compensate the problem that the image size affects thesegmentation in local threshold method. Finally dynamic threshold for imagesegmentation is adopted and the integrity and accuracy of defect extraction areimproved.(5) In the classification of extraction apparent defect unit, the local binary pattern(LBP) is combined with image variance intensity feature and then the operator LBPC isproposed. By chi-square formula, the feature classification distance of defect sampletraining and test sets are calculated, thus completing the automatic classification ofdefect types. The experimental results for the accuracy of apparent defect classification have improved significantly. Proposed algorithm has been combined with traditionaladaptive neural network classification algorithm, the classification accuracy has beenincreased by12%to95.5%, which meets the need of industion(6) In the acceleration of algorithm system unit, the graphics processing unit (GPU)is studied and the device effect of real-time is improved. Meanwhile the efficiency ofparallel processing of the CUDA design pattern and applied the technique in thecomplex algorithms improvement of PCB automatic optical detection systems arein-depth analyzed. The experimental results have indicated that the system achievesparallel processing acceleration for the apparent defect image preprocessing, extractionand automatic classification algorithm, which greatly reduces the running time of thesystem. Computing speed can increase30times for large amount of date in imageprocessing.In this thesis, through research on PCB apparent defects automated opticalinspection theory and key technology, we proposed the apparent defect detection systemdesign scheme, image processing algorithms and classification methods. And then thecomplexity algorithm system has been reduced by using computer graphics and imageprocessing unit, improved the execution efficiency of image processing and thereal-time of apparent automatic detection system. The average detection time is only3swhen completion entire board of25cm×22cm PCB. Finally, the apparent automaticdetection system has been verified successfully in the actual projects and put intoindustrial application.更多还原