【作者】 毕昕；

【导师】 丁汉；

【作者基本信息】 上海交通大学 ， 机械电子工程， 2009， 博士

【摘要】 随着液晶显示器(LCD)向大尺寸、轻薄化、低功耗、高分辨率的方向发展,玻璃基板及相关光学组件的尺寸逐渐增大、其厚度日益减小,致使液晶显示器产生各种显示不均匀缺陷(Mura)的几率大大增加,传统的人眼缺陷检测方法受人的主观因素以及外界环境的影响严重,对缺陷等级没有统一的判断标准,且由于LCD大尺寸高分辨率的趋势,使人工检测方法很难满足产品质量和生产效率的要求。因此,研究快速、不受外界环境干扰、符合人眼判断标准的自动机器视觉缺陷检测方法成为发展液晶显示技术的迫切要求。本文面向TFT-LCD生产制程,立足缺陷检测这一关键问题,开展Mura缺陷机器视觉检测方法的研究。Mura缺陷对比度低,边缘模糊或无明显边缘,图像背景复杂,是最难检测的显示缺陷之一,传统的边缘检测以及阈值分割等方法很难实现缺陷的可靠检测。本文针对Mura缺陷的检测难点,围绕纹理背景抑制、亮度不均匀校正、缺陷分割和缺陷量化四个关键点,通过理论研究、仿真分析和实验验证,研究符合人眼判断标准的自动机器视觉检测方法。主要工作及取得的成果如下:一、针对Mura缺陷图像具有重复纹理背景的特点,提出了基于实值Gabor小波滤波的纹理背景抑制方法,并研究了滤波器参数的设计准则。根据实际纹理特点和检测目的而设计的实值Gabor小波滤波器,可作为Blob检测器,能很好的消除纹理背景并增强待检测缺陷。与现有的常用背景抑制方法相比,Gabor小波滤波方法与人眼的视觉特性相似,能与人工检测的结果更好的吻合;能适应在线检测中图像背景不均匀以及旋转变形等复杂情况,具有较好的鲁棒性。二、在分析了LCD图像背景亮度不均匀产生原因的基础上,将各种因素引起的亮度不均统一看作乘性照度不均,采用同态变换的方法将乘性照度偏差场转换为加性,然后应用独立分量分析的方法对线性混合信号进行估计和分离,从而分离出缺陷信号,消除亮度不均匀信号。通过仿真分析,验证了所提出的方法可以在不掌握图像亮度先验信息的情况下,较好的消除图像背景的整体不均匀以及由于频率混叠而引起的莫尔条纹。三、在对LCD图像进行纹理背景抑制和亮度不均匀校正的基础上,研究Mura缺陷的分割方法。针对Mura缺陷对比度低、边缘模糊、图像整体亮度不均匀的特点,通过比较现有的基于边界和基于区域的主动轮廓模型的优缺点,采用改进的Chan-Vese主动轮廓模型和水平集方法对缺陷进行自动分割。与传统的边缘检测方法相比,改进的Chan-Vese主动轮廓模型能捕捉缺陷的模糊边界,可以减小图像整体亮度不均匀对Mura缺陷分割造成的影响。同时,根据SEMI标准中关于Mura缺陷的等级评定方法,直接应用分割时所得到的缺陷对比度和面积参数,对Mura缺陷进行量化评定。四、根据SEMI标准中关于Mura缺陷的测量规范,基于LCD缺陷检测平台,采集带有Mura缺陷的LCD图像作为实验样本,分别对LCD图像纹理背景抑制、亮度校正、Mura缺陷的分割及量化方法进行实验验证。(1)通过实验比较,分析了滤波器各个参数对纹理抑制效果的影响和参数选用原则,验证了所设计的实值Gabor小波滤波器能快速的消除LCD的纹理背景,有较好的鲁棒性,抑制效果符合人眼判断标准。(2)验证了与传统分割方法相比,改进的Chan-Vese主动轮廓模型和水平集方法适用于Mura缺陷分割。采用无条件稳定的AOS格式求解水平集方程可以增大时间步长、节省求解时间。同时,把分割中求出的缺陷对比度和面积参数直接应用到缺陷的量化中,方便地得出了Mura缺陷样本的量化实验结果。(3)验证了所提方法能较好的校正LCD亮度不均匀,通过比较实验,表明经过亮度校正,图像分割更易实现,分割结果更准确。(4)建立了Mura缺陷的自动检测算法流程并进行实验验证,对于50个缺陷样本,有48个被成功检测。更多还原

【Abstract】 With the development of LCD toward large area, thin thickness and high resolution, the size of glass substrate increases rapidly and the thickness decreases gradually, which makes a remarkable occurrence probability of Mura defect. Due to high resolution, the traditional manual inspection method hardly satisfies the quality and efficiency of LCD manufactures, thus a fast and objective automatic machine vision inspection way that in accordance with the human criterion is very important to LCD insdustry.This thesis makes studies of Mura defect inspection technique based on the key problem of vision inspection for TFT-LCD. Mura is local lightness variation with low contrast, blurry contour and complicated image background, so it is hard to be inspected with traditional thresholding or edge detection methods. This thesis is to study the automatic vision inspection way for Mura defect which in accordance with the human criterion, focusing on textural background suppression, uneven brightness adjustment, defect segmentation and quantification. The main contents and achievement are as follows:Firstly, considering the respective textural background of Mura defect, a background suppression method based on the real Gabor wavelet filter is proposed and the design principles of filters are studied. Designed according to the texture characteristics, the real Gabor wavelet filter can eliminate the background and enhance Mura defect, acting as an excellent blob detector. Comparing with usual background suppressin method, Gabor wavelet filter is similar with the vision characteristic of human eyes and can obtain good accordance with manual inspection results. And it’s robustness to adapt to the backgroud uneveness, image rotation and distortion.Secondly, through analyzing the reasons of uneven brightness, all uneven factors are treated as the multiplicative illuminance. The homomorphic filtering method is used to transform the multiplicative uneveness into additive one, and then the independent component analysis method is applied to the estimation and separation of mixed signals. Thus the uneven brightness signals are eliminated or lightened and the target defective singals are preserved. The simulation results show that the proposed method can realize blind separation of uneven brightness and moire fringe.Thridly, the segmentation method of Mura defet is studied after background suppression and brightness adjustment. By comparing the edge-based and region-based active contour models, a modified Chan-Vese active contour model together with the level set method is proposed to automatically segment the Mura defect, aiming at the characteristics of low contrast, blurry contour and background uneveness of Mura. The modified model can exactly trace the blurry contours and eliminate the influences of uneven background luminance. At the same time, the parameters of Mura contrast and area from segmentation are directly used in the defect quantification based on SEMI standard.Finally, the image samples of LCD with Mura defects are captured based on the regulations of SEMI standard, as well as, the background suppression,brightness adujstment and Mura segmentation method proposed above are verified. For real Gabor wavelet filter, the influences of parameters on the background suppression effects are compared and the design principles are proved through experiments. Results show that this method can eliminate the textural background of LCD rapidly and reliably. Further, for the modified Chan-Vese active contours model together with the level set method, it is appropriate and excellent to the Mura segmentaion. The unconditional stable AOS scheme can increase the time step and save the solution time of level set function. Then, with the parameters of Mura contrast and area from segmentation, the quantifiction experiments are carried out conveniently. In addition, the proposed method can realize good brightness adjustment of LCD images and the defect segmentation becomes easier and more exactly after adjustment. Finally, the automatic inspection process of Mura defect is set up. Inspection experiments show that for 50 samples with Mura, 48 samples are inspeced accurately.更多还原