| 摘要: | 高密度印刷電路板（PCB）製程中的瑕疵檢測對產品品質至關重要，然而現有自動視覺檢查（AVI）系統在面對細微或不規則缺陷時仍常需依賴人工覆判，影響檢測效率與一致性。本研究為解決此問題，提出一套基於深度學習的兩階段缺陷檢測與分類架構，並設計以實務應用為導向，易於整合於產線工作流程中。本系統首先透過分類模型進行二元缺陷檢測，進一步結合物件偵測技術執行多類別缺陷分類，藉此降低人工負擔，同時兼顧高靈敏度與分類準確性。
在缺陷檢測階段，本研究建構三組資料集，分別對應三種不同PCB產品，並依生產標準進行標註。比較ResNet-50與ResNeXt-101兩種卷積神經網路架構後，ResNeXt-101於獨立的2,141張實際產線測試集中展現優異表現，召回率達97.98%，逃脫率僅1.63%，整體準確率為96.92%。為實現實地應用，本研究以Tkinter開發圖形化使用介面，使用CSV資料格式操作，便於產線人員進行批次檢測與結果匯出，無需具備深度學習背景。
在缺陷分類階段，本研究統整NG樣本構建統一資料集，針對四類常見缺陷進行框選與標註，並分別訓練Faster R-CNN與Cascade R-CNN模型。實驗結果顯示Cascade R-CNN整體表現較佳，平均召回率達93.20%，逃脫率為1.65%，在微小與形狀不規則之缺陷上具有更佳辨識能力，亦能偵測未標註之可疑區域，顯示其已學習可泛化之缺陷特徵。模型在定位精度與小目標偵測方面的優勢亦透過可視化結果進行驗證。
綜上所述，本研究所提出之框架能在多樣化PCB情境下維持高準確率與低逃脫率，並已成功導入實際產線應用，展現其於現代電子製造業中智慧化缺陷檢測之潛力與實用價值。;Defect inspection in high-density printed circuit board (PCB) manufacturing is critical to ensuring product reliability, but existing automated visual inspection systems often require additional manual verification due to limitations in detecting subtle or irregular defects. To address these challenges, this study proposes a deep learning-based two-stage framework for automated defect detection and classification, designed for integration into practical industrial workflows. The system integrates a classification-based binary defect detection stage and an object-detection-based defect classification stage, aiming to reduce reliance on manual inspection while maintaining high detection sensitivity and classification accuracy.
In the defect detection stage, three real-world PCB datasets, each corresponding to a specific product type, were constructed and labeled according to production line standards. Two CNN architectures, ResNet-50 and ResNeXt-101, were employed and evaluated using five-fold cross-validation. ResNeXt-101 demonstrated strong real-world performance when evaluated on an independent test set composed of 2,141 PCB images collected from an actual production line, achieving a recall of 97.98%, an escape rate of only 1.63%, and an overall accuracy of 96.92%. To facilitate integration into the production environment, a Tkinter-based graphical user interface (GUI) was developed. The interface supports CSV-based batch inference, allowing production-line operators to perform automated inspections and receive classification results without requiring technical expertise.
For the defect classification stage, a unified dataset of NG samples was constructed, containing four dominant defect types annotated with bounding boxes. Faster R-CNN and Cascade R-CNN were trained and evaluated using five-fold cross-validation. Cascade R-CNN achieved a higher average recall of 93.20%, lower escape rate of 1.65%, and better performance on subtle or irregular defect types. It also detected previously unlabeled defect-like regions, demonstrating the model’s ability to generalize beyond explicit training annotations. Visualization results confirmed its advantages in localization precision and small object detection.
The results demonstrate that the proposed framework achieves high accuracy, low escape rates, and strong generalization performance across diverse PCB conditions. Its deployment-ready design, validated in a real production environment, highlights its practical applicability for intelligent defect inspection in modern electronics manufacturing. |
