集成电路芯片表面缺陷视觉检测关键技术研究

发布时间：2018-05-26 08:48

本文选题：集成电路芯片 + 缺陷检测　；参考：《东南大学》2016年博士论文

【摘要】：集成电路芯片已广泛应用于多个领域,但是制造过程中产生的缺陷会直接影响集成电路芯片的寿命和可靠性。传统的人工检测方法,存在耗时长,劳动强度大,误检率高等缺点,已无法适应生产的需求。机器视觉检测技术通过机器视觉的方法对产品进行分析处理,检验产品是否符合质量要求,对保障产品质量,提高产品合格率起到了关键作用。本文结合国家自然科学基金资助项目(50805023)和江苏省科技成果转化专项(BA2010093),以集成电路芯片表面缺陷为研究对象,展开视觉检测关键技术研究,所从事的主要研究工作如下:(1)集成电路芯片表面缺陷图像多阈值分割针对集成电路芯片表面缺陷图像的特点,提出基于萤火虫算法的二维熵多阈值缺陷图像分割法。首先,将二维熵阈值分割扩展为二维熵多阈值分割。其次,分析萤火虫算法仿生原理和寻优过程。最后,将二维熵作为萤火虫算法的目标函数,对多阈值寻优。实验结果表明,基于萤火虫算法的二维熵多阈值缺陷图像分割法能有效分割集成电路芯片表面缺陷;运行速度较穷举法有很大的提高;在阈值选取的准确性、计算时间和峰值信噪比方面均优于基于粒子群算法的二维熵多阈值分割法;但是仍然存在实时性不足的问题。针对集成电路芯片表面缺陷图像多阈值分割计算量大、计算时间长的问题,提出基于反向萤火虫算法的大津多阈值缺陷图像分割法。首先,将大津阈值分割扩展为大津多阈值分割。其次,提出一种反向萤火虫算法。该算法将反向学习算法中反向解可能比当前解距离目标函数更近的思想引入萤火虫算法,增加萤火虫的多样性和全局搜索能力。最后,将最大类间方差作为反向萤火虫算法的目标函数,对多阈值寻优。实验结果表明,基于反向萤火虫算法的大津多阈值缺陷图像分割法的性能优于穷举法、基于粒子群算法、萤火虫算法的大津多阈值分割法:但是该分割法在四阈值分割时出现了一些寻优结果不准确的现象。为了分割集成电路芯片表面缺陷图像,提出基于改进的萤火虫算法的大津多阈值缺陷图像分割法。针对萤火虫算法全局搜索和局部搜索不平衡的现象,提出改进的萤火虫算法。在该算法中,提出基于Cauchy变异的多样性增强策略和邻域策略,并根据不同的停滞状态,选择不同的策略以增强全局搜索能力并提高收敛性能。将改进的萤火虫算法应用于大津多阈值分割中,对阈值寻优。实验结果表明,基于改进的萤火虫算法的大津多阈值缺陷图像分割法不仅能有效分割缺陷图像,并在准确性、计算时间、收敛性和稳定性方面整体优于基于达尔文粒子群算法、混合的差分进化算法、萤火虫算法、反向萤火虫算法的大津多阈值分割法。(2)集成电路芯片表面缺陷提取集成电路芯片表面缺陷明场图像中存在噪声干扰缺陷的提取。为了提取明场图像中的缺陷,提出基于数学形态学变换和改进的区域生长的缺陷提取法。首先,根据图像灰度级,获得明场图像四阈值分割后的浅层缺陷图像和深层缺陷图像。其次,针对浅层缺陷图像和深层缺陷图像的不同特点,将不同的数学形态学操作应用于两种图像以去除噪声点,定位缺陷所在的区域。最后,提出改进的区域生长法来提取明场图像中的缺陷。实验结果表明,该方法能有效提取明场图像中的缺陷。集成电路芯片表面缺陷暗场图像中存在芯片表面纹理干扰缺陷的提取。为了提取暗场图像中的缺陷,提出基于纹理方向检测和缺陷区域选择的缺陷提取法。首先,针对暗场图像的特点,提出芯片表面纹理方向检测算法。其次,根据图像灰度级,获得暗场图像四阈值分割后的暗缺陷图像和明缺陷图像。最后,针对暗缺陷图像和明缺陷图像的不同特点,以缺陷纹理方向与芯片表面纹理方向不一致为原则,提出两种不同的缺陷区域选择的方法以提取暗场缺陷。实验结果表明,该方法能有效提取暗场图像中的缺陷。(3)集成电路芯片表面缺陷特征提取与降维为了提取集成电路芯片表面缺陷特征,分别从几何特征、纹理特征和灰度特征三个方面提取32个特征。提取的几何特征包括面积、周长、紧凑性、重心坐标、矩形度、占空比、偏心率和Hu不变矩。提取的纹理特征为14个灰度共生矩阵特征。提取的灰度特征包括灰度均值、灰度方差和灰度熵。由于特征维数较多,采用主成分分析特征抽取法和基于KNN的序列浮动前向特征选择法分别进行特征降维。主成分分析法根据特征值累积贡献率的取值大于90%的原则,将32维特征降至6维。基于KNN的序列浮动前向特征选择法将每个特征的KNN分类性能作为序列浮动前向选择的评价函.数以实现特征选择,将32维特征降至10维。(4)集成电路芯片表面缺陷分类识别为了识别并分类集成电路芯片表面缺陷,分析讨论了 BP神经网络和RBF神经网络。为了提高支持向量机的分类识别率,提出基于改进的萤火虫算法的支持向量机,其基本思想是将分类识别率作为目标函数,通过改进的萤火虫算法对支持向量机中的惩罚参数和核函数参数进行寻优。将8种芯片缺陷对应的主成分分析法抽取的6维特征和基于KNN的序列浮动前向特征选择法选择的10维特征分别输入三个分类器,形成六组分类器。实验结果表明,特征选择法选择的特征作为基于改进的萤火虫算法的支持向量机的输入时,分类性能高于其他五组分类器,识别率为91.367%。本文对缺陷多阈值分割、缺陷提取、缺陷特征提取与降维、缺陷分类识别等视觉检测关键技术进行研究,在理论研究和技术研发等方面取得了一定的成果,为集成电路芯片表面缺陷视觉检测提供了理论指导和技术支撑。
[Abstract]:Integrated circuit chips have been widely used in many fields, but the defects produced in the manufacturing process will directly affect the life and reliability of integrated circuit chips. The traditional artificial detection methods have the disadvantages of long time consuming, high labor intensity and high false detection rate, which have been unable to meet the needs of production. Machine vision detection technology is used by machine vision. Methods to analyze and deal with the product, test whether the product meets the quality requirements, and play a key role in ensuring the quality of the product and improving the rate of product qualification. This paper combines the National Natural Science Fund Project (50805023) and the Jiangsu province science and technology achievement transformation special (BA2010093), and the integrated circuit chip surface defect as the research object. The main research of the key technology of visual detection is as follows: (1) a two-dimensional entropy multi threshold defect image segmentation method based on the firefly algorithm is proposed. First, the two-dimensional entropy threshold segmentation is extended to a two-dimensional entropy multithreshold. Secondly, the bionic principle and optimization process of the firefly algorithm are analyzed. Finally, the two-dimensional entropy is used as the target function of the firefly algorithm to optimize the multi threshold. The experimental results show that the two-dimensional entropy multi threshold defect image segmentation method based on the firefly algorithm can effectively segment the surface defects of the integrated circuit chip, and the running speed is more than the poor method. Large improvement, the accuracy of threshold selection, calculation time and peak signal to noise ratio are superior to two-dimensional entropy multi threshold segmentation based on particle swarm optimization, but there is still a problem of lack of real time. First, the Otsu threshold segmentation is extended to the multi threshold segmentation of the Otsu threshold. Secondly, a reverse firefly algorithm is proposed. The algorithm introduces the idea that the reverse solution in the reverse learning algorithm may be more close to the target function than the current solution to the firefly algorithm to increase the diversity of the firefly. In the end, the maximum inter class variance is used as the target function of the reverse firefly algorithm to optimize the multi threshold. The experimental results show that the performance of the multi threshold image segmentation method based on the reverse firefly algorithm is superior to the exhaustive method, based on the particle swarm optimization and the multi threshold segmentation method of the firefly algorithm, but the segmentation method is used. In order to divide the image of the surface defect of the integrated circuit chip, a new method of multi threshold defect image segmentation based on improved firefly algorithm is proposed in order to divide the surface defect image of the integrated circuit chip. The improved firefly algorithm is proposed for the global search and local search imbalance of the firefly algorithm. In the algorithm, the diversity enhancement strategy and neighborhood strategy based on Cauchy mutation are proposed, and different strategies are selected to enhance the global search ability and improve the convergence performance according to the different stagnation states. The improved firefly algorithm is applied to the Otsu multi threshold segmentation and the threshold optimization. The experimental results show that the improved firefly calculation is based on the improved firefly calculation. The method not only can effectively segment the defect image, but also is better than the Darwin particle swarm optimization, the hybrid differential evolution algorithm, the firefly algorithm, the reverse firefly algorithm and the Otsu multi threshold segmentation method. (2) the surface deficiency of the integrated circuit chip. In order to extract the defects in the image of the bright field, a defect extraction method based on the mathematical morphology transformation and the improved region growth is proposed. First, the shallow defect image and the deep defect image after the four threshold segmentation of the field image are obtained according to the gray level of the image. Secondly, in view of the different characteristics of the shallow defect image and the deep defect image, different mathematical morphology operations are applied to two images to remove the noise point and locate the region where the defect is located. Finally, an improved region growth method is proposed to extract the defect in the bright field image. The experimental results show that the method can effectively extract the bright field. The defect in the image of the integrated circuit chip is extracted from the defect in the surface texture of the chip. In order to extract the defects in the dark field image, a defect extraction method based on the texture direction detection and the defect region selection is proposed. Firstly, the texture direction detection algorithm of the chip surface is proposed for the characteristics of the dark field image. At the same time, the dark defect image and the bright defect image after the four threshold segmentation are obtained according to the gray level of the image. Finally, according to the different characteristics of the dark defect image and the bright defect image, two different defect region selection methods are proposed to extract the dark field defect with the principle that the defect texture direction is inconsistent with the texture direction of the chip surface. The experimental results show that the method can effectively extract the defects in the dark field image. (3) the feature extraction and dimension reduction of the surface defects of the IC chip are extracted and reduced to extract the features of the surface defects of the integrated circuit chip, and 32 features are extracted from the geometric features, the texture features and the gray features respectively. The extraction geometry features include the area, the circumference, and the compact features. Sex, center of gravity, rectangles, duty ratio, eccentricity, and Hu invariant moments. The extracted texture features are 14 gray level co-occurrence matrix features. The extracted gray level features include gray mean, gray variance and gray entropy. The feature extraction method and KNN based sequence floating forward feature selection method are used respectively because of the large number of feature dimensions. The principle component analysis is based on the principle that the value of the cumulative contribution rate of the eigenvalue is greater than 90%, and reduces the 32 Vitter sign to 6 dimension. The KNN classification performance of each feature is used as the evaluation function of the sequence floating forward selection based on the sequence floating forward feature selection method based on KNN. The feature selection is realized and the 32 Vitter sign is reduced to 10 dimension. (4) integration In order to identify and classify the surface defects of IC chips, the BP neural network and RBF neural network are analyzed and classified. In order to improve the classification recognition rate of support vector machines, a support vector machine based on improved firefly algorithm is proposed. The basic thought is to use the classification recognition rate as the target function. The improved firefly algorithm optimized the penalty parameters and kernel parameters in the support vector machine. The 6 dimension feature extracted by the principal component analysis and the 10 dimension feature selection based on the KNN based sequence floating forward feature selection method, respectively, input three classifiers to form six groups of classifiers. The experimental results show that the characteristics of the 8 kinds of chip defects are characterized. The feature of selection method selected as the input of support vector machine based on improved firefly algorithm, the classification performance is higher than the other five groups of classifier. The recognition rate is 91.367%.. This paper studies the key technology of defect multi threshold segmentation, defect extraction, defect feature extraction and dimensionality reduction, defect classification recognition and so on. Some achievements have been achieved in technology research and development, providing theoretical guidance and technical support for visual inspection of integrated circuit chip surface defects.
【学位授予单位】：东南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN407;TP391.41

【相似文献】