A Novel Self-Feedback Intelligent Vision Measure for Fast and Accurate Alignment in Flip-Chip Packaging
A Novel Self-Feedback Intelligent Vision Measure for Fast and Accurate Alignment in Flip-Chip Packaging
复制标题
DOI：
10.1109/tii.2019.2930078
复制
发表时间：
2020-03
影响因子：
12.3
通讯作者：
Zelong Wu;Hui Tang;Zhaoyang Feng;Weimin Wang;Sifeng He;Jian Gao;Xin Chen;Yunbo He;Xun Chen
中科院分区：
文献类型：
--
作者：
Zelong Wu;Hui Tang;Zhaoyang Feng;Weimin Wang;Sifeng He;Jian Gao;Xin Chen;Yunbo He;Xun Chen
模板匹配（TM）算法已广泛应用于晶圆级倒装芯片封装的视觉检测过程中，但传统的TM算法结构多为直接前馈，难以同时实现快速和高精度。本文的动机是联合收割机的能力，使芯片的视觉测量运行在一个快速和高精度的方式。首先，提出了一种新的自反馈智能模板匹配（SFI-TM）结构，它可以使匹配过程中的中间信息得到充分利用。结合智能速度和分辨率调节规则，构造SFI-TM算法。然后，从理论上分析了SFI-TM算法的可靠性和鲁棒性，以确保其以稳定的方式工作。最后，进行了一系列实际的芯片对准视觉检测实验，包括参数测试、与其他五种视觉检测算法的对比实验以及鲁棒性测试。实验结果表明，SFI-TM算法可以达到最高的平均测量精度（1.08 $\mu$m），测量速度与Tiny YOLOv 2几乎相同，并且可以抵抗从$-$20到30灰度值的亮度变化、密度为0.5/pixel的椒盐噪声和方差为2.5的大高斯噪声。此外，它对动态速度运动过程中图像的模糊和失真具有良好的鲁棒性。
A template matching (TM) algorithm has been widely employed in a visual inspection process of wafer-level flip-chip packaging, but the structures of traditional TM algorithms are always direct feedforward, which leads to the difficulty in achieving fast speed and high accuracy at the same time. The motivation of this article is to combine the ability to enable the chip visual measurement running in a fast-speed and high-accuracy manner. First, a novel self-feedback intelligent template matching (SFI-TM) structure is proposed, which can enable the intermediate information in the matching process to be fully utilized. The intelligent speed and resolution regulation rules are combined to construct the SFI-TM algorithm. Then, the reliability and robustness of the SFI-TM algorithm are theoretically analyzed to make sure it works in a stable manner. Finally, a series of practical chip alignment visual inspection experiments, including parameter testing, comparing experiments with the other five proposed visual detection algorithms, and robustness testing, are carried out in detail, respectively. The experimental results indicate that the SFI-TM algorithm can achieve the highest average measurement accuracy (1.08 $\mu$m) with almost the same measurement speed as fast as Tiny YOLOv2, and that it can resist the brightness variation from $-$20 to 30 gray value, the pepper–salt noise with a density of 0.5/pixel, and the Gaussian noise with a large variance of 2.5. In addition, it has good robustness against blurring and distortion of images under the dynamic speed motion processes.