Ball bonding inspections with human vision are essential in manufacturing processes of semiconductors devices and integrated circuits (ICs). The inspections are an intensive task which involves human labours to detect poor bonds. Prolonged visual inspections cause poor inspection integrity due to eye-fatigue. However, inspections nowadays are mostly conducted manually by humans which cannot satisfy the demanding productions. Motivated by the extraordinary performance of machine learning for manufacturing inspections, a detection framework integrated with machine learning and human judgement is proposed to aid bonding inspections based on visual images. The detection framework is incorporated with the convolution neural network (CNN), support vector machine (SVM) and circle hough transform algorithm (CHT); human judgement is only used when the detection uncertainty is below the threshold. The novel machine learning integration is proposed on the detection framework to improve the generalization capabilities. The CNN architecture is redeveloped by incorporating with the SVM which is generally more effective than the fully connected network in the classical CNN. Also a novel training function is proposed based on the CHT to ensure the inspection reliability; the function not only takes into account real image captures, but also locates important features using pattern analysis of the ball bondings. Experimental results show that significantly better classifications can be achieved by the proposed framework compared with the classical CNN and other commonly used classifiers. Only the machine learning determinations below the threshold are reassessed by human judgements.

在半导体器件和集成电路（IC）的制造过程中，具有人类视觉的球焊检查至关重要。检查是一项艰巨的任务，需要人工来检测不良的结合。长时间的目视检查会由于眼睛疲劳而导致检查完整性差。但是，如今的检查大多是由人为进行的，无法满足生产要求。出于机器学习在制造检查中非凡的性能的激励，提出了一种与机器学习和人为判断相集成的检测框架，以帮助基于视觉图像的粘合检查。该检测框架与卷积神经网络（CNN），支持向量机（SVM）和圆霍夫变换算法（CHT）结合在一起；仅当检测不确定性低于阈值时才使用人工判断。在检测框架上提出了新颖的机器学习集成，以提高泛化能力。通过与SVM合并来重新开发CNN架构，该SVM通常比传统CNN中的全连接网络更有效。为了保证检验的可靠性，还提出了一种基于CHT的新型训练功能。该功能不仅考虑了真实的图像捕获，而且还使用了球键合的图案分析来定位重要特征。实验结果表明，与经典的CNN和其他常用分类器相比，该框架可以实现更好的分类。