Research on PCBA defect detection system based on computer vision
Cheng, Muqin (2025)
Kandidaatintyö
2025
School of Engineering Science, Tietotekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2025052855514
https://urn.fi/URN:NBN:fi-fe2025052855514
Tiivistelmä
Defects on PCBAs have become common in electronics manufacturing, posing challenges to automated inspection due to their small size, irregular shapes, and blurred boundaries. Traditional methods often fail to meet industrial requirements for accuracy and reliability.
To address this issue, two image datasets were used in this study: a custom-constructed dataset focusing on non-structural defects, and a public dataset containing five types of structural anomalies. Building upon this dataset, an enhanced detection framework based on YOLOv8 was proposed. Key improvements include module-level architectural refinements to strengthen multi-scale feature aggregation and optimized training configurations to boost detection accuracy and robustness.
The improved model achieved 92.4% Precision, 92.3% Recall, and 90.0% mAP@0.5 on the non-structural dataset, demonstrating significant gains over the baseline model. On the structural defect dataset, the optimized model kept excellent performance with 99.5% Precision, and 70.2% mAP@0.5:0.95, though the enhancements over the baseline were less significant.
This work shows the benefit of model optimization in boosting small object detection of surface-level defects while maintaining the robustness of the performance in structured anomaly cases. Future work could try to address the limitations by increasing number and types of defects, and enriching diversity of dataset, and improving the application efficiency in real-world applications.
To address this issue, two image datasets were used in this study: a custom-constructed dataset focusing on non-structural defects, and a public dataset containing five types of structural anomalies. Building upon this dataset, an enhanced detection framework based on YOLOv8 was proposed. Key improvements include module-level architectural refinements to strengthen multi-scale feature aggregation and optimized training configurations to boost detection accuracy and robustness.
The improved model achieved 92.4% Precision, 92.3% Recall, and 90.0% mAP@0.5 on the non-structural dataset, demonstrating significant gains over the baseline model. On the structural defect dataset, the optimized model kept excellent performance with 99.5% Precision, and 70.2% mAP@0.5:0.95, though the enhancements over the baseline were less significant.
This work shows the benefit of model optimization in boosting small object detection of surface-level defects while maintaining the robustness of the performance in structured anomaly cases. Future work could try to address the limitations by increasing number and types of defects, and enriching diversity of dataset, and improving the application efficiency in real-world applications.