Automatic optical inspection is a vital part of the production process for solder joints appearance inspection in surface mounted technology assembling lines. However, IC solder joint inspection is a challenging problem because IC solder joints have extremely small sizes and no distinct appearance differences between qualified and unqualified ones. In this paper, we propose an IC solder joint inspection method via generator-adversarial-network based template. We are the first to introduce the GAN strategy into IC solder joint inspection. The method consists of GAN template generator training, offline statistical modelling and online real-time inspection. At the training stage, the GAN template generator is trained based on a designed GAN, which involves the feature maps in both of high-dimension and low-dimension spaces. Then, the binary difference image can be achieved by the input IC solder joint image and the corresponding GAN-based template. At the offline statistical modelling stage, to reduce the interferences, a pixel probability image is statistically modelled by the binary difference images corresponding to qualified IC solder joints. At the online real-time inspection stage, the potential defect pixels for the inspected IC solder joint can be shown in a defect salient image achieved by the multiplication of its corresponding binary difference image and the pixel probability image. Finally, we can accumulate the pixels in the defect salient image to distinguish the quality of the inspected IC solder joint. Experimental results show that the proposed method is superior to the state-of-the-art inspection methods with 0% omission rate and 0.15% error rate at a reasonable inspection speed of 4.32 ms per sample.

自动光学检测是表面贴装技术组装线焊点外观检测生产过程的重要组成部分。然而，IC焊点检测是一个具有挑战性的问题，因为IC焊点尺寸极小，合格和不合格之间没有明显的外观差异。在本文中，我们提出了一种基于生成器-对抗网络模板的 IC 焊点检测方法。我们是第一个将 GAN 策略引入 IC 焊点检测的公司。该方法由GAN模板生成器训练、离线统计建模和在线实时检测组成。在训练阶段，GAN 模板生成器基于设计的 GAN 进行训练，其中涉及高维和低维空间中的特征图。然后，二值差分图像可以通过输入的 IC 焊点图像和相应的基于 GAN 的模板来实现。在离线统计建模阶段，为了减少干扰，通过合格IC焊点对应的二值差分图像对像素概率图像进行统计建模。在在线实时检测阶段，被检测IC焊点的潜在缺陷像素可以显示在缺陷显着图像中，通过其对应的二进制差异图像和像素概率图像相乘得到。最后，我们可以累积缺陷显着图像中的像素来区分被检测IC焊点的质量。实验结果表明，所提出的方法优于最先进的检查方法，遗漏率为 0% 和 0.