面向IC封装设备的图像匹配研究

发布时间：2019-01-06 13:14

【摘要】：随着信息技术的发展，IC(Integrated Circuit)技术在国防、航天航空、通信等各行各业及日常生活中得到了广泛应用。机器视觉技术作为IC封装设备的关键共性技术，伴随着IC封装产业的发展而逐渐走向成熟。由于研发比较早、技术比较成熟，国外的商用机器视觉软件包占据了绝大部分市场份额，如MIL、HALCON和Congex等。这些视觉软件包使用方便、性能稳定、效率较高，但是费用昂贵。因此，研发拥有自主知识产权的视觉软件包具有重大的意义。本文对视觉软件包中的图像匹配算法进行了研究，分别提出了基于Zernike矩的灰度匹配算法和基于广义Hough变换的特征匹配算法，以满足IC封装设备中对定位技术高速度、高精度、适应旋转的应用要求。图像匹配算法可以分为两个大类，基于灰度的匹配算法和基于特征的匹配算法。基于灰度的匹配方法对噪声具有较好的鲁棒性，与特征匹配相比较，其能够适用模糊图像的匹配。基于特征的匹配方法对光照有较强的适应性，且能够对残缺、遮挡等情况进行匹配。基于灰度和基于特征的匹配算法各有其优缺点，在应用中应该结合实际情况选择合适的匹配方法。在基于灰度匹配算法中，提出了基于Zernike矩的快速匹配算法。首先根据三角函数的对称性，仅需要计算八分之一个圆内的Zernike矩基函数值，快速计算Zernike矩；然后通过求取图像多个矩的互相关性，确定匹配点的像素位置；进而通过最小二乘拟合，得到亚像素位置；最后利用Zernike矩的相位信息，估算旋转角度。同时，采用离线阶段建立RCS表，保存与快速计算Zernike矩的相关信息，在线阶段查找RCS表的方法进一步加速匹配算法。实验结果表明，该算法位置匹配精度可以达到0.5个像素内，角度匹配精度可以达到0.1°以上，快速算法匹配速度为非快速算法的2~3倍。在基于特征匹配算法中，提出了适应旋转的广义Hough变换匹配算法。该算法在广义Hough变换思想的基础上，利用梯度方向与旋转角度之间的关系，实现对存在旋转的图像进行投票匹配。并对投票模型进行了分析，提出了高斯投票模型。在理论分析的基础上，设计实现了该算法。首先在离线阶段建立Rtable；然后在在线阶段查找Rtable，对匹配位置和匹配角度进行投票，得到位置和角度；最后利用最小二乘拟合，得到亚像素位置和角度。通过测试结果可知，该算法能够实现对旋转图像进行匹配，位置匹配精度可以达到0.5个像素内，角度匹配精度可以达到0.5°以上，，零角度匹配时间为16ms，全角度匹配时间为70ms。
[Abstract]:With the development of information technology, IC (Integrated Circuit) technology has been widely used in national defense, aerospace, communications and other industries and daily life. Machine vision technology, as the key common technology of IC packaging equipment, has gradually matured with the development of IC packaging industry. Because of the early R & D and mature technology, most of the foreign software packages, such as MIL,HALCON and Congex, occupy most of the market share. These visual software packages are easy to use, stable, efficient, but expensive. Therefore, it is of great significance to develop visual software packages with independent intellectual property rights. In this paper, the image matching algorithm in visual software package is studied, and the gray level matching algorithm based on Zernike moment and the feature matching algorithm based on generalized Hough transform are proposed, respectively, to satisfy the high speed and high precision of positioning technology in IC packaging equipment. Adapt to the application of rotation requirements. Image matching algorithms can be divided into two categories: gray-based matching algorithm and feature-based matching algorithm. The method based on gray level is robust to noise. Compared with feature matching, it can be used to match fuzzy images. The feature-based matching method has strong adaptability to illumination and can match incomplete occlusion and other cases. The matching algorithm based on gray level and feature has its own advantages and disadvantages, so the suitable matching method should be selected according to the actual situation in the application. In the grayscale matching algorithm, a fast matching algorithm based on Zernike moment is proposed. Firstly, according to the symmetry of trigonometric function, the Zernike moment basis function in an eighth circle is only calculated, and the Zernike moment is calculated quickly, then the pixel position of the matching point is determined by calculating the mutual correlation of several moments in the image. Then the sub-pixel position is obtained by least square fitting. Finally, the rotation angle is estimated by using the phase information of Zernike moment. At the same time, the RCS table is set up in the off-line stage to save and quickly calculate the relevant information of the Zernike moment, and the method of searching the RCS table in the online phase further accelerates the matching algorithm. The experimental results show that the accuracy of position matching can reach 0.5 pixels, the precision of angle matching can reach 0.1 掳, and the matching speed of the fast algorithm is 2 times than that of the non-fast algorithm. In the feature matching algorithm, the generalized Hough transform matching algorithm adapted to rotation is proposed. Based on the idea of generalized Hough transform and the relationship between gradient direction and rotation angle, the algorithm realizes the voting matching of images with rotation. The voting model is analyzed and Gao Si voting model is put forward. On the basis of theoretical analysis, the algorithm is designed and implemented. Firstly, the Rtable; is established in the off-line phase, and then the Rtable, is found in the online stage to vote on the matching position and the matching angle. Finally, the location and angle of the sub-pixel are obtained by least square fitting. The test results show that the algorithm can match rotating images, the accuracy of position matching can reach 0.5 pixels, the precision of angle matching can reach 0.5 掳, the time of zero-angle matching is 16ms. The full angle matching time is 70 Ms.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】