| 摘要: | 本研究針對工業4.0環境下的晶片瑕疵檢測需求，提出一個結合物件偵測、對比學習與知識蒸餾的AI檢測系統，旨在提升多產品場景下的瑕疵辨識效能與模型泛化能力。針對半導體封裝製程中常見的微小瑕疵與產品特徵差異，傳統影像檢測方法容易因資料不足或模型過度擬合導致準確度下降。為解決此問題，本研究設計並實作三大技術模組： 首先，從 Waffle 中精準切出各區域晶片位置，確保輸入影像具備區域性特徵，利於後續分析晶片間空間關係與整體排列。接著，採用對比學習模型，除學習單顆晶片好壞特徵外，亦利用多晶片影像的相對位置關係，偵測晶片偏移或對位異常，提升異常檢測能力。最後，導入知識蒸餾 (Knowledge Distillation, KD) 技術，讓不同產品類別進行互相學習，透過教師-學生架構將跨產品特徵整合至單一模型中，提升多產品場景下的準確率與通用性。
本研究使用多組不同產品的晶片影像資料集進行模型訓練與測試，並從切圖準確率、分類準確率 (Accuracy)、精確率 (Precision)、召回率 (Recall)、特異度(Specificity)等多面向評估效能。實驗結果顯示，YMK-FDC架構在多產品場景下表現優於單一模型與傳統監督式學習方法。多產品知識蒸餾技術使模型即使面對未見類別亦具備良好辨識效果，對比式學習對多晶片區域之比對則提升對偏移異常之敏感度，展現其在多樣性資料中的學習潛力與適應性。本研究不僅提供創新整合方案，亦具高度實務應用價值，可廣泛應用於晶圓、封裝基板等半導體製程瑕疵檢測場景，對提升智慧製造產線的自動化檢測能力具重要意義。所建立的多產品學習架構與知識蒸餾模型，亦為未來多模態學習與跨產品研究提供關鍵參考。;This study addresses the demand for chip defect inspection in Industry 4.0 environments by proposing an AI-powered detection system that integrates object detection, contrastive learning, and knowledge distillation to enhance defect recognition performance and model generalization across multiple product types. Traditional image-based inspection methods often suffer from accuracy degradation due to limited data or model overfitting when confronted with the tiny defects and inter-product feature variability commonly found in semiconductor packaging processes. To overcome these challenges, we design and implement three core modules:
First, precise chip localization is performed on the Waffle substrate to ensure that each input image contains well-defined regional features, facilitating subsequent analysis of spatial relations and overall chip arrangement. Next, a contrastive learning module not only learns the good-vs-defective characteristics of individual chips but also exploits the relative positions of chips in multi-chip images to detect displacement or misalignment abnormalities, thereby improving anomaly detection capability. Finally, we introduce Knowledge Distillation (KD) to enable cross-product learning: through a teacher–student framework, features learned from one product category are transferred and integrated into a single student model, boosting accuracy and robustness in mixed-product scenarios.
We train and evaluate our system on multiple chip image datasets spanning different product lines, measuring performance across region-cropping accuracy, classification accuracy, precision, recall, and specificity. Experimental results demonstrate that the YMK-FDC architecture outperforms single-model baselines and conventional supervised learning approaches in multi-product settings. The multi-product KD technique allows the student model to generalize effectively even to unseen classes, while the contrastive module’s focus on inter-chip region relationships enhances sensitivity to alignment anomalies—highlighting the approach’s adaptability and learning potential in diverse datasets.
This work not only offers an innovative, integrated solution with strong practical value—applicable to defect inspection in wafer and substrate packaging stages of semiconductor manufacturing—but also lays a foundation for future research in multimodal learning and cross-product generalization. |
