【作者】 黄丹；

【导师】 胡跃明；

【作者基本信息】 华南理工大学 ， 控制科学与工程， 2021， 博士

【摘要】 超薄高密度柔性集成电路封装基板(FICS)在柔性电路板基础之上,往更高集成密度、更轻薄、更易弯曲扭转的方向发展。随着IC行业进入7-14纳米制程,超薄高密度FICS也进入2m以下制程,对材料和制造过程的品质控制性能要求也不断提升。在FICS表面缺陷检测过程中,仅靠基板传统检测方法,已无法达到工业生产的精度要求。在FICS图像预处理环节中,存在去除噪音的同时去除原有图像的纹理信息等问题。对于高倍显微镜下得到的FICS基板显微成像的图像,其纹理结构和缺陷同时被放大,各类缺陷呈现多样性,其中氧化区域的像素值呈现离散型分布,存在氧化缺陷检测不精准等问题。对于无参考模板的超薄高密度FICS线路检测,检测结果容易出现线路不连续且线路缺陷定位不精准等问题。如何设计出一种同时具有高精度、高效率且鲁棒性强的检测算法成为一大难点。本文主要对超薄高密度FICS外观视觉检测中出现的问题,提出以下五种数学模型:1、在超薄高密度FICS图像预处理环节中,提出两种结合水平集曲率特征与梯度特征的平滑模型。为了有效去除噪音的同时,保留更多的图像原有细节。首先,本文提出的两个模型将图像的水平集曲率特征与梯度阈值相结合,使用更丰富的二阶微分信息作为检测因子去除图像中的噪声。其次,理论分析表明,所提出的两个模型获得的去噪图像可以保留原始图像更详细的纹理信息和边缘信息。此外,实验分析表明,与其他模型相比,所提出的模型具有最高的结构相似性和峰值信噪比,并且具有相对较高的边缘保持指数和最低的均方误差。具体地,利用模型一去噪后图像具有最高的结构相似性和峰值信噪比,以及最低的均方误差。利用模型二去噪图像具有相对较高的边缘保持指数。本文提出的方法可以有效去除超薄高密度FICS图像的噪声,并保留图像的原始细节和边缘信息。2、在超薄高密度FICS图像预处理环节中,提出一种融合梯度和自适应曲率特征的图像去噪模型。本文所提出的模型能够自适应地调节图像的水平集曲率特征与梯度特征所占的权重,将更丰富的图像一阶微分与二阶微分信息作为检测因子进行图像去噪处理。理论分析表明:本文模型在图像平坦区域中的扩散力度大于PM模型;在图像的角点、尖峰处,本文模型可以抑制灰度值的减小,保留图像更多细节及边缘信息。实验分析表明:将本文模型分别与P-M模型、C模型、G-C模型进行去除高斯噪音和椒盐噪声实验对比,本文的模型均有最高的结构相似度、峰值信噪比及边缘保持指数。本文所提出的方法能够有效地去除超薄高密度柔性FICS图像的噪音,同时更多的保留图像原有的细节信息及边缘信息,具有实际工程意义。3、针对高倍镜下的超薄高密度FICS表面氧化缺陷分布检测不精准问题,提出一种区域型变分水平集图像分割模型。在超薄高密度FICS图像缺陷检测环节中,首先,对于演化曲线内部的图像,本文利用K-means聚类得到的聚类中心点的均值与滤波后的图像的加权和,作为演化曲线内部的拟合值;同时,对于演化曲线内部的图像滤波处理,提出一种融合水平集曲率与梯度特征的各向异性扩散模型进行滤波;其次,对于演化曲线外部的图像,引进图像熵作为灰度信息的补充;其次,利用最大绝对均值差,自适应调节演化曲线内外权重比。最后,将该算法应用于超薄高密度FICS表面氧化缺陷检测中,实验对比结果表明,应用本文所提出的区域型变分水平集分割方法,使得提取后的氧化缺陷具有更高的分割精度。4、针对无参考模板的超薄高密度FICS线检测中线路不连续,且缺陷定位不准确的问题,提出了一种提取FICS图像线路特征的检测算法。在超薄高密度FICS图像缺陷检测环节中,本文所提出的检测算法步骤:首先将彩色FICS图像进行K-Means分类;其次进行中值滤波、形态学填充、闭运算,从而得到待分割线路图像的二值图像;最后一步,提出一种保凸间接正则水平集的图像分割模型,应用在FICS二值图像中,从而达到提取超薄高密度FICS图像线路特征的目的。所提出的能量泛函模型由数据项、连接项、正则项组成。数据项引导轮廓向目标边界运动;连接项是水平集函数与辅助函数之间的连接项,使得水平集函数在演化过程中不能偏离辅助函数;正则项是辅助函数的正则项,避免水平集函数在演化过程中出现震荡,并且弱化了水平集函数的过度平滑,避免边界泄露。本文所提出的模型在给定L~2(Ω)空间上为严格凸函数,存在全局唯一极小值。实验仿真表明:将本文的算法应用在提取FICS线路特征中,能够实现线路特征的精准提取,并且提取的线路边界平滑,为高精度测量线宽线距以及高精度定位缺陷位置做出重要铺垫。更多还原

【Abstract】 Ultra-thin and high-density flexible integrated circuit packaging substrates(FICS)are developing in the direction of higher integration density,lighter and thinner,more flexible and twisted based on flexible circuit boards.As the IC industry enters the 7-14nanometer process,ultra-thin and high-density FICS has also entered the process below2m,and the quality control performance requirements for materials and manufacturing processes are also constantly improving.In the FICS surface defect detection process,only relying on the traditional detection methods of substrates can no longer meet the accuracy requirements of industrial production.In FICS image preprocessing,there are problems such as removing noise while removing texture information of the original image.For the FICS substrate microscopic imaging image obtained under a high-power microscope,the texture structure and defects are magnified at the same time,and various types of defects show diversity.Among them,the pixel value of the oxidation area shows a discrete distribution,and there are problems such as inaccurate detection of oxidation defects.For ultra-thin and high-density FICS line inspection without reference template,the inspection results are prone to cause problems such as line discontinuity and inaccurate positioning of line defects.How to design a high-precision,efficient,and robust detection algorithm at the same time has become a major problem.This paper mainly proposes the following five mathematical models for the problems that occur in the appearance of ultra-thin and high-density FICS visual inspection:1.In the pre-processing of ultra-thin and high-density FICS images,two smooth models that combine level set curvature features and gradient features are proposed.In order to effectively remove noise while retaining more original image details.First,the two models proposed in this paper combine the level set curvature feature of the image with the gradient threshold,and use richer second-order differential information as a detection factor to remove noise in the image.Secondly,theoretical analysis shows that the de-noised image obtained by the two proposed models can retain more detailed texture information and edge information of the original image.In addition,experimental analysis shows that compared with other models,the proposed model has the highest structural similarity and peak signal-to-noise ratio,and has a relatively high edge retention index and the lowest mean square error.In particular,the de-noised image through model 1has the highest structural similarity and peak signal-to-noise ratio,and the lowest mean square error.The de-noised image through model 2 has a relatively high edge preservation index.The method proposed in this paper can effectively remove the noise of ultra-thin and high-density FICS images,and retain the original details and edge information of the image.2.In the preprocessing of ultra-thin and high-density FICS images,an image de-noising model that combines gradient and adaptive curvature features is proposed.The model proposed in this paper can adaptively adjust the weight of the level set curvature and gradient features of the image.Through this method,richer image first-order and second-order differential information can be used as detection factors for image de-noising processing.Theoretical analysis shows that:the diffusion power of the model in the flat area of the image is greater than that of the P-M model;at the corners and peaks of the image,the model in this paper can suppress the reduction of gray value and retain more details and edge information of the image.The experimental analysis shows that the model in this paper is compared with the P-M model,C model,and G-C model to remove Gaussian noise and salt and pepper noise.Experimental results show that the model in this paper has the highest structural similarity,peak signal-to-noise ratio and edge retention index.The method proposed in this paper can effectively remove the noise of ultra-thin,high-density flexible FICS images,while retaining more original details and edge information of the image,which has practical engineering significance.3.Aiming at solving the problem of inaccurate detection of surface oxidation defect distribution in ultra-thin and high-density fisc under high magnification,a regional vari-ational level set image segmentation model is proposed.In the process of ultra-thin and high-density FICS image defect detection,first of all,for the images inside the evolution curve,this paper uses the weighted sum of the mean value of the cluster center points obtained after K-means clustering and the filtered image as the fitting value inside the evolution curve.At the same time,for the image filtering inside the evolution curve,an anisotropic diffusion model that combines the curvature and gradient features of the level set is proposed for filtering.Secondly,the model in this paper uses the maximum absolute mean difference to adaptively adjust the weight ratio between the inside and outside of the evolution curve.Finally,the algorithm is applied to the detection of ultra-thin and high-density FICS surface oxidation defects.The experimental comparison results show that the application of the regional variational level set segmentation method proposed in this paper makes the extracted oxidation defects have higher segmentation accuracy.4.Aiming at solving the problem of discontinuous lines and inaccurate defect loca-tion in the ultra-thin and high-density FICS line inspection without reference template,a detection algorithm for extracting FICS image line features is proposed.In the process of ultra-thin and high-density FICS image defect detection,firstly,the detection algo-rithm steps proposed in this paper perform K-Means classification of color FICS images.Secondly,perform median filtering,morphological filling,and closing operations to ob-tain the binary image to be segmented.In the last step,an image segmentation model with convexity-preserving and indirect regularity level set is proposed,which is used in extracting ultra-thin and high-density FICS image line features.The proposed energy functional model consists of data item,connection item,and regular item.The data item guides the contour to move to the target boundary;the connection item is the connection item between the level set function and the auxiliary function,so that the level set func-tion cannot deviate from the auxiliary function in the evolution process;the regular item is the regular item of the auxiliary function,which prevents the level set function from oscillating during the evolution process,and the excessive smoothness of the level set function is weakened to avoid boundary leakage.The model proposed in this paper is a strictly convex function in a given L~2(Ω)space,and there is a globally unique minimum.Experimental simulation proves that the application of the algorithm in this paper in extracting FICS line features can achieve precise extraction of line features and smooth line boundaries,which is an important pavement for high-precision measurement of line width and line spacing,and also is an important pavement for high-precision location of defects.更多还原