面向极大规模集成电路封装X射线检测的图像处理关键问题研究

发布时间：2018-11-12 14:19

【摘要】：极大规模集成电路产业是国民经济和国防建设的重要支柱之一,是衡量一个国家和地区科技实力的重要标志。随着集成电路多功能、高密度、集成化多层封装的发展,传统的自动光学检测只能对表面缺陷进行检测,难以实现集成化多层封装的内部缺陷检测。X射线通过透射成像可直观地检测出元器件的内部缺陷,为极大规模集成电路产业提供了非常有效的检测手段。然而,集成电路多由高密度材质组成且结构细小,形成的X射线图像具有信噪比低、对比度低、有效特征细小等特点,难以实现有效的检测。内部缺陷检测精度直接影响集成电路产业的封装质量,因此,研究面向集成电路封装X射线检测的图像复原方法具有理论意义和很高的应用价值。本文面向极大规模集成电路X射线检测研究了正则化框架下的图像复原问题,重点研究了适用于稀疏性约束的2 1l-l模型和适用于噪声去除的全变分(Total Variation,TV)正则化模型下的快速性算法以及在物理运动角度和导数空间上的新模型建立。主要研究工作如下:1)研究了一种基于动量梯度下降方向的快速去模糊算法,解决了常规梯度下降方向在处理2 1l-l模型下的图像复原问题时收敛速度慢、对噪声极为敏感的问题。该算法通过模拟物理学中的动量性质,引入了动量梯度下降方向作为迭代方向,使得算法在平滑区域增大步长,提高算法收敛速度;在噪声区域利用当前迭代的动量特性阻止下降速度的突然改变,避免算法收敛至局部最优解。本文在理论上给出并证明了算法稳定解存在的充要条件。标准灰度图像和集成电路X射线图像两个系列的实验证明了算法的快速性和有效性。2)通过近点理论推导得到了一类正则化项不可微的复原模型在动量梯度下降方向下的凸二阶近似模型,结合全变分正则化项提出了基于动量梯度投影的TV正则化图像复原算法并在理论上证明了其收敛性,避免了TV正则项的不可微性带来的计算困难。最后,实验结果证明了该算法在计算速度上的优越性以及在去噪和细节保持方面的有效性。3)从物理运动角度出发,将算法迭代过程看作粒子的移动过程,在2 1l-l模型基础上提出了基于物理总能量目标函数的稀疏重建模型及算法,并对算法的收敛性进行了证明,解决了传统方法中模型非单调非强凸且建模及求解过程均未考虑物理演化规律的问题。其中,粒子的运动模型是在粘性介质、牛顿流体中以迭代结果为位移、迭代次数为离散时间、2 1l-l模型为重力势能函数建立的运动模型。与多种算法的对比实验证明了本文方法速度快、重建质量高等优点,表明了该方法更适合于实际的X射线缺陷检测。4)利用图像空间和导数空间的关联性,在导数空间中建立了基于梯度分离的各向异性TV和各项同性TV正则化图像复原模型,并采用Split Bregman框架提出了相应的各向异性算法和各向同性算法,解决了传统模型保真项未考虑导数空间具有提高图像复原成功率这一优势的问题。该方法在导数空间内实现了水平梯度和垂直梯度的解耦,有利于图像去噪和细节保持能力的提升。实验结果表明,相较于传统图像空间中的TV正则化方法,该方法具有更优的去噪能力和细节保持特性。
[Abstract]:The very large-scale integrated circuit industry is one of the important pillars of national economy and national defense construction, and is an important sign to measure the strength of science and technology in a country and region. With the development of the multi-function, high-density and integrated multi-layer package of the integrated circuit, the traditional automatic optical detection can only detect the surface defects, and it is difficult to realize the internal defect detection of the integrated multi-layer package. The X-ray can directly detect the internal defect of the component through the transmission imaging, and provides a very effective means for detecting the large-scale integrated circuit industry. however, that integrate circuit is composed of high-density material and the structure is small, the formed X-ray image has the characteristics of low signal-to-noise ratio, low contrast, small effective characteristic and the like, and is difficult to realize effective detection. The precision of internal defect detection directly affects the packaging quality of the integrated circuit industry. Therefore, the research of the image restoration method facing the integrated circuit package X-ray detection is of great theoretical significance and high application value. In this paper, the problem of image restoration under the regularized framework is studied for the X-ray detection of a very large-scale integrated circuit. The 2l-l model and the total variation for noise removal are studied. The fast algorithm under the regularization model of TV and the establishment of a new model on the physical motion angle and the derivative space. The main research work is as follows: 1) A fast de-fuzzy algorithm based on momentum gradient descent direction is studied, which solves the problem that the convergence speed is slow and the noise is extremely sensitive when the image restoration problem of the conventional gradient descent direction under the processing of the 2l-l model is solved. The method introduces the momentum gradient descent direction as the iteration direction by simulating the momentum property in physics, so that the algorithm increases the step size in the smooth region, and the convergence speed of the algorithm is improved; and the sudden change of the falling speed is prevented by the momentum characteristic of the current iteration in the noise region, and the algorithm is avoided to converge to the local optimal solution. In this paper, the necessary and sufficient conditions for the existence of the stable solution of the algorithm are given in this paper. The experiments of two series of standard gray-scale image and integrated circuit Xray image show the rapidity and validity of the algorithm. 2) The convex second-order approximate model of a kind of regularized non-differentiable recovery model in the direction of momentum gradient descent is derived by the near-point theory. In this paper, a TV regularization image restoration algorithm based on momentum gradient projection is proposed in combination with all-variable regularization term, and its convergence is proved theoretically, and the computational difficulty of the non-differentiability of the TV regular term is avoided. In the end, the experimental results show the superiority of the algorithm in the calculation speed and the validity of the de-noising and detail maintenance. 3) From the point of physical motion, the iterative process of the algorithm is regarded as the moving process of the particles. On the basis of the 2l-l model, a sparse reconstruction model and an algorithm based on the physical total energy objective function are proposed, and the convergence of the algorithm is proved. The motion model of the particle is in the viscous medium, the iteration result is the displacement in the Newtonian fluid, the number of iterations is the discrete time, and the 2l-l model is the motion model established by the gravity potential energy function. The comparison experiments with a variety of algorithms have proved that the method has the advantages of high speed, high reconstruction quality and the like, and shows that the method is more suitable for practical Xray defect detection. 4) The relevance of the image space and the derivative space is utilized, In the derivative space, an anisotropic TV based on gradient separation and an image restoration model of the same-sex TV are established, and a corresponding anisotropic algorithm and an isotropic algorithm are proposed by using the Split Bregman framework. The problem that the traditional model fidelity term does not take into account the derivative space has the advantage of improving the image restoration success rate is solved. The method realizes the decoupling of the horizontal gradient and the vertical gradient in the derivative space, and is beneficial to the improvement of the image de-noising and the detail maintenance capability. The experimental results show that the method has better de-noising capability and detail-preserving characteristics than the TV regularization method in the traditional image space.
【学位授予单位】：华南理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】