Deep Learning-based AOI System for Detecting Component Marks

While the automated optical inspection (AOI) was adopted for the surface mounted technology (SMT) on printing circuit board assembly (PCBA), it inspected that the products with miniaturized components and high-density placement of components on boards are becoming important issues. Especially, due to the manufacturing trend in the high-mix and low volume, the detection of the component marks with wide variation are growing into even more challengeable. As for the defect detection of AOI machine by using the similarity technology, it is very sensitive to different marks, including the different brightness of the font, spacing, brand, word shift and direction. Therefore, it implies that the AOI performance which is reduced as the AOI engineer fails to solve the demanding problem, and results in more time on checking the alarmed images by a human operator. Also, it affects the production yield rate and overall equipment effectiveness. Consequently, in this paper, a novel architecture is proposed to solve such a problem. That is, the classification method combined with similarity measurement on basis of the deep learning structure. The technique we used takes the MobileNet lightweight model to modify as the output with double stages to prevent the escape from occurring. The proposed method is approved by the manufacturing line to produce the products due to its no escape occurred, easier model training and faster implement. The proposed method provides a high improved rate of the average 90% with the zero-leakage rate in SMT process according to the experimental results.