【作者】 陈安；

【导师】 胡跃明；

【作者基本信息】 华南理工大学 ， 控制理论与控制工程， 2013， 博士

【摘要】 柔性印制电路板行业经过六十多年的发展，在新材料、新工艺的带动下，不断和上下游产业相互促进甚至融合发展，已经成为国防军工，芯片制造等具有战略意义行业中不可或缺的支柱产业。本文研究工作的主要目的为提高柔板行业的自动化水平，研制自动化设备，提高生产效率，降低废品率。本文通过对柔板行业现有生产方式、工艺环节的分析，总结出制约该行业进一步提高自动化水平的四个典型工艺环节。并深入分析这些环节中的实际问题，抽象总结成为两类理论问题：图像配准问题和表面缺陷检测问题。通过对比人类和现有机器视觉算法在解决这两类问题时的不同方法和表现，揭示出人类利用选择性注意机制和信息逐层抽取方式以较少的数据量完成了图像配准和缺陷检测的事实。并设计了四个人类实验，以实验数据证明了相关理论和猜想。本文从仿生学角度出发，尝试模拟人类在解决图像配准和表面缺陷检测这两类问题时所采用的视觉机制。为了达到此目的，首先提出了创新性的“A-B”并行采集处理模式及其机构模型。在这个机构模型的基础上，以显著性计算为基础，模拟人类的选择性注意机制和信息抽取方式，把图像配准问题和表面缺陷检测问题整合到一个框架里。为了讨论显著性，也为了解决柔板图像的干扰边缘问题，本文首先讨论了主边缘提取方法。由于尺度空间的许多算子和人类视网膜以及视觉皮层的感受野剖面有着很高的相似度，本文先讨论了基于尺度空间的边缘提取。并提出了一种新的基于尺度空间和格式塔率模拟的主边缘提取法。该方法能较好的去除保护膜、X光图像纹理造成的干扰边缘，得到主要对象的清晰轮廓，同时也不影响缺陷的外观。本文重点讨论了显著性的生成。首先讨论了基于频域的显著性计算方法，介绍了现有方法，并提出了使用唯相位变换来进行计算的新方法。进而讨论了图像金字塔里该显著性方法的特点。即低分辨率下的频域显著性图较好的模拟了选择性注意机制，而高分辨率下的频域显著性图凸显了包括缺陷在内的非常规模式，可以作为很好的缺陷检测手段。其次讨论了基于对象的显著性计算方法，并提出了使用色彩先验知识来指导对象分割合并的新方法。第三提出了新的显著性概念，即基于缺陷的显著性计算方法。该显著性计算方法模拟人类自上而下的串行视觉处理过程产生高层知识指导下的显著性图。基于缺陷的显著性图一共有三种：色彩缺陷显著性图，边缘密度显著性图和边缘规整度显著性图。在讨论了显著性图的特性后，本文提出了一种基于显著性计算和局部不变特征的图像配准方法。主要讨论了使用频域显著性方法和最大稳定极值区域法相结合的图像配准方法。该方法使用图像金字塔中的低分辨率显著性图作为数据导引，在显著性点周围进行开窗操作得到显著性区域。而最大稳定极值区域法就在所开的显著性区域窗口中进行操作。通过这样的处理方法，减少了总体计算量，降低了误匹配的可能性。并通过显著性的导引形成了簇状的数据空间，加速了匹配速度。为了进行对比，也讨论了使用基于对象显著性方法和局部不变特征法相结合的图像配准方法。最后本文提出了一种基于显著性计算和信息熵倍率的表面缺陷检测方法，这种方法分为两个阶段。第一阶段工作在“A-B”模式下，当两幅图像粗配准完成后，将两幅图像的高分辨率显著性点集进行几何变换后的双向比较，进而得出缺陷候选对象。而第二阶段使用唯相位变换前后的信息熵倍率来专门处理焊盘金面的缺陷检测问题。这两阶段方法都是对人类缺陷检测行为的模拟，而模拟的关键也是显著性计算方法。更多还原

【Abstract】 Flexible Printed Circuit (FPC) board industry has developed for more than60years.Nowadays, led by the new material and new technology, FPC becomes more and more criticalin national industrial system. Especially when chip manufacturing industry uses FPC as ICsubstrate, it becomes a strategic pillar industry. The main purpose of the research work in thispaper is to improve the automation level of the FPC industry and improve productionefficiency, reduce defects rate as well.This paper analyses the cutting-edge technology and equipment used in FPC industry.And summaries four bottleneck processes for improving automation. The problems in thesefour processes are abstracted into two basic theoretical issues—the image registration andsurface defects inspection problems. By comparing the human and the existing machinevision algorithm in solving these two problems, this paper reveals that human uses selectiveattention mechanism and information extraction step by step way to reduce the amount of datafor visual task. And we design four human experiments to prove the related theory andhypothesis.Inspired by human vision system, we try to solve these two problems by using selectiveattention and effective information extraction. First of all, we construct mechanic model andpropose novel―A-B‖parallel mode. In this mode, there are two cameras capture a picture andb picture at once. And this two pictures are designed to imaged roughly the same parts of twoproducts. In this way, the image gets reference for each other and doubles the throughput.Saliency map is used to simulate human selective attention. And by this concept we put imageregistration and surface defects detection problem into one computational framework.For saliency computation, we propose a main edge abstraction method. Scale spaceanalysis is used to abstract main edges in a FPC image. And Gestalt theory are simulated toimprove the main edge map. This method is very good in abstract main edge in FPC imageand X-ray image and has not affection to defect image.We discuss three kinds of saliency computation. The first one is saliency map based onfrequency domain method. We propose using phase-only Fourier transform to calculatefrequency domain saliency map. The different resolution frequency domain saliency maps arediscussed to illustrated how to used them as index of special signals. The second one issaliency map based on object and color. We propose a novel method based on objectsegmentation-merge and object priori knowledge. The third one is saliency map based ondefect feature. We propose three defect saliency maps: the color defect saliency map, the edge density saliency map and edge regularity saliency map.After discussing of saliency map, we propose a novel image registration method base onsaliency man and MSERs. The low resolution saliency map is used for selective attentionclues and sub windows are abstracted around the saliency extreme points. Then the MSERsare searched in the sub windows. By this way, we reduce the amount of the feature andmismatching possibility.After registration of two―A-B‖mode images，we transform the local extreme points ofhigh resolution saliency map to inspect the candidates of defects. And the defect of platedgold surface is discussed specially. We propose a novel method based on saliency map andShannon entropy ratio to detect defects on plated gold surface. And this method is alsosimulation of human vision.更多还原