半导体芯片封装过程中视觉定位关键技术研究

发布时间：2018-03-18 21:57

本文选题：芯片封装　切入点：视觉定位　出处：《华中科技大学》2016年博士论文　论文类型：学位论文

【摘要】：在半导体芯片封装过程中,视觉定位技术是实现芯片精确拾取和转移的核心技术之一。近年来出现的细长芯片和超薄高密度芯片对现有封装技术提出了新的要求。为了解决以往视觉定位算法在新的封装需求下定位不准确或无法定位的问题,本文提出了基于多级教与学优化算法的相机标定方法、有效区域归一化互相关匹配算法和空间结构约束特征点匹配算法,现将本文研究内容总结如下：1)提出了基于多级教与学优化算法的相机标定方法。多级教与学优化算法是对教与学优化算法的改进,本文创新性地引入了“校长”的概念,提出了“老师”之间相互学习的思想。相比于原始的只有“老师”和“学生”两个角色的教与学算法,多级教与学优化算法具有更强的全局搜索能力,在精确求取相机内部参数、外部参数和畸变系数时更容易跳出局部最优解,提高了相机标定精度和芯片封装过程中的定位精度。2)提出了有效区域归一化互相关匹配算法RBNCC.该算法通过区域置信函数对图像中的芯片和背景像素进行分类加权,解决了现有灰度算法无法精确定位细长芯片和有效识别破损芯片的问题,提高了细长芯片封装过程中的定位精度和良品率。提出了芯片边界框快速检测方法,可将图像搜索范围缩减到芯片周围区域,加快了RBNCC算法的定位速度,提升了半导体封装装备效率。3)提出了空间结构约束特征点匹配算法SCFP。超薄高密度芯片电路图案复杂,特征丰富,因此非常适合使用特征点匹配算法进行定位,但是现有算法无法在同一幅图像中匹配多颗芯片。为了解决该问题,SCFP算法不仅利用了特征点周围的局部信息,还利用了特征点之间的拓扑关系,从而可以将目标图像中的特征点有效地划分到不同芯片区域。结合本文提出的基于多级教与学优化算法的相机标定方法,推导了芯片三维姿态计算公式,为多自由度拾取头准确拾取超薄高密度芯片、防止其破碎提供了依据。4)为了缩短视觉定位算法的计算时间,提升芯片封装效率,本文分析了有效区域归一化互相关匹配算法和空间结构约束特征点匹配算法中可并行执行的任务,充分利用GPU在并行计算、内存管理等方面的优势,合理分配了主机端和设备端的资源以及在两种算法中承担的角色,实现了两种算法的加速。5)构建了具有自主知识产权的图像处理函数库MV1.0,定义了基于多级教与学优化算法的相机标定方法、有效区域归一化互相关匹配算法和空间结构约束特征点匹配算法的用户接口函数,完成了三种算法在MV1.0中的集成,设计了智能工具软件MVPlat,该软件界面友好,操作简便,方便进行算法验证。最后在典型半导体封装装备上进行了应用与测试,测试结果表明三种算法满足实际应用需求,具有广泛的应用前景。
[Abstract]:In the process of semiconductor chip packaging, Visual positioning technology is one of the core technologies to realize the accurate pickup and transfer of chips. In recent years, thin and long chips and ultra-thin high density chips have put forward new requirements for the existing packaging technology. In order to solve the problem of visual positioning calculation in the past. To locate problems that are inaccurate or unable to locate under the new packaging requirements, This paper presents a camera calibration method based on multi-level teaching and learning optimization algorithm, an effective region normalized cross-correlation matching algorithm and a spatial structure constrained feature point matching algorithm. In this paper, the following is summarized as follows: 1) A camera calibration method based on multi-level teaching and learning optimization algorithm is proposed. The multi-level teaching and learning optimization algorithm is an improvement on the teaching and learning optimization algorithm. In this paper, the concept of "principal" is introduced innovatively. This paper puts forward the idea of "teachers" learning from each other. Compared with the original teaching and learning algorithms with only the roles of "teacher" and "student", the multi-level teaching and learning optimization algorithm has a stronger global search ability. It is easier to jump out of the local optimal solution when the camera internal parameters, external parameters and distortion coefficients are accurately obtained. An effective region normalized cross-correlation matching algorithm RBNCCS is proposed, which classifies and weights the chip and background pixels in the image by using the region confidence function (RBC), which improves the calibration accuracy of the camera and the positioning accuracy in the chip packaging process. It solves the problem that the existing gray level algorithm can not accurately locate the thin chip and effectively identify the damaged chip, and improves the positioning accuracy and the good product rate in the process of packaging the slender chip. A fast detection method for the chip boundary frame is proposed. The search range of image can be reduced to the area around the chip, the localization speed of the RBNCC algorithm is accelerated, and the efficiency of semiconductor packaging equipment is improved. Because of the rich features, it is very suitable to use the feature point matching algorithm for localization, but the existing algorithms can not match multiple chips in the same image. In order to solve this problem, the SCFP algorithm not only makes use of the local information around the feature points. The topological relationship between the feature points is also used to divide the feature points into different chip regions effectively. The camera calibration method based on the multi-level teaching and learning optimization algorithm is proposed in this paper. The three-dimensional attitude calculation formula of the chip is derived, which provides a basis for accurately picking up ultra-thin and high-density chips with multi-degree-of-freedom pick-up heads and preventing them from breaking) in order to shorten the computing time of the visual positioning algorithm and improve the packaging efficiency of the chip. In this paper, we analyze the tasks that can be performed in the efficient region normalized cross-correlation matching algorithm and the spatial structure constraint feature point matching algorithm, and make full use of the advantages of GPU in parallel computing, memory management, etc. Reasonable allocation of host and device resources and roles in the two algorithms, The image processing function library MV1.0 with independent intellectual property rights is constructed, and the camera calibration method based on multi-level teaching and learning optimization algorithm is defined. The user interface function of the effective region normalized cross-correlation matching algorithm and the spatial structure constraint feature point matching algorithm is implemented. The integration of the three algorithms in MV1.0 is completed. The intelligent tool software MVPlatis is designed. The interface of the software is friendly and the operation is simple. Finally, it is applied and tested on typical semiconductor packaging equipment. The test results show that the three algorithms meet the needs of practical applications and have a wide application prospect.
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】