應用機器學習於 PCB 瑕疵檢測

Abstract

本研究針對印刷電路板（Printed Circuit Board, PCB）瑕疵檢測中的準確性與效率問題，提出了一種基於深度學習模型 YOLOv8 的改進方法。隨著電子產品日益精密化，PCB 製造過程中出現的各類缺陷，包括 Missing Hole、Mouse Bite、Open Circuit、Short、Spur 和Spurious Copper，成為影響產品品質的主要挑戰。傳統的人工檢測與自動光學檢測（Automated Optical Inspection, AOI）系統存在準確性與效率的限制，促使本研究探索深度學習技術的應用。本研究設計的實驗架構，包含資料擴增策略、模型架構設計、損失函數改進與訓練策略。資料擴增過程中，對影像進行旋轉、縮放、亮度與對比度調整，產生多樣化的訓練樣本。模型選擇了 YOLOv8作為基礎，並針對其邊界框定位損失函數進行改進，依次測試了CIoU、MSE、MAE、 Huber Loss 與中心點距離損失等多種函數，尋求最佳配置。實驗採用了 K-fold 交叉驗證法，以評估不同參數組合的模型性能。結果顯示，MSE 與 Huber Loss 在模型性能上表現最佳，特別是在精度與召回率的平衡上具有顯著優勢。此外，針對新瑕疵類別的引入，本研究提出了兩種訓練策略：整合式訓練與單類別訓練。以調整新瑕疵類別的損失權重與數據分佈，成功提高了模型在少數樣本場景下的檢測準確性。綜合實驗結果顯示，經改進的 YOLOv8 模型在 PCB 瑕疵檢測中的精度和穩健性均達到了令人滿意的水準。

This study addresses the challenges of accuracy and efficiency in printed circuit board (PCB) defect detection by proposing an improved method based on the deep learning model YOLOv8. As electronic products become increasingly sophisticated, various defects such as missing holes, mouse bites, open circuits, shorts, spurs, and spurious copper frequently occur during PCB manufacturing, posing significant quality assurance challenges. Traditional manual inspections and automated optical inspection (AOI) systems have limitations in terms of accuracy and efficiency, motivating this research to explore deep learning applications.The proposed experimental framework includes data augmentation strategies, model architecture design, loss function improvements, and training strategies. During data augmentation, images were processed with rotations, scaling, brightness, and contrast adjustments to generate diverse training samples. The model selected YOLOv8 as the foundation and improved its bounding box localization loss function by testing various alternatives such as CIoU, MSE, MAE, Huber Loss, and Center Point Distance Loss to identify the best configuration. A K-fold cross-validation approach was used to evaluate model performance under different parameter combinations.Results indicate that MSE and Huber Loss demonstrated the best model performance, particularly in balancing precision and recall. Additionally, the study proposed two training strategies for introducing new defect types: integrated training and single-class training. Adjusting the loss weights and data distribution of the new defect class successfully improved detection accuracy in limited-sample scenarios. Comprehensive experimental results show that the improved YOLOv8 model achieved satisfactory accuracy and robustness in PCB defect detection.