| dc.description.abstract | With the development of hardware technology, the demand for circuit boards has increased dramatically year by year, and the industry′s demand for quality and yield of circuit boards has also been increasing. In particular, the rapid growth in hardware performance over the past few years has allowed deep learning technology to advance by leaps and bounds. Deep learning technology is often used in everyday life and industry, for example, GOOGLE translation and car park license plate recognition.
At present, PCB yield inspection in the PCB printing-related industry mainly relies on automated optical inspection systems (AOI) and manual inspection, which requires a large amount of manual inspection due to the frequent error in defect judgement in AOI systems, resulting in higher costs.
This study proposes a deep learning-based inspection technique to detect defects on PCBs, with the main objective of filtering out ′false defects′ marked by AOI systems. After filtering by the deep learning neural network system developed in this study, the workload and labour cost can be significantly reduced and the efficiency and yield of PCB defect detection can be significantly improved.
In this study, the main architecture of the neural network system is YOLO, which is a very powerful neural network system for object detection in recent years, widely used in academia and industry. It has been trained to accurately determine and mark the location of targets.
The dataset consists of images from the AOI system of the production line provided by the manufacturer, which contains images of 11 types of defects and non-defects that were misidentified by the AOI inspection system.This study is different from other experiments in terms of difficulty.
The test data in this study is the current image output from the AOI system of PCB production line sent by the vendor every month. Compared to other experiments, the test data is closed data with similar characteristics, and usually a part of the database is divided into test data, which is extremely difficult.
| en_US |