基於深度學習的自動光學檢查方法在最近幾年快速取代傳統的方法。如何提高瑕疵檢測的準確率是自動光學檢查重要的環節。集成學習是應用多重的機器學習模型來提高準確率。如何集成不同的深度學習模型來提高準確率也是自動光學檢查的重要課題。 在本論文中，先使用遷移式學習訓練8個基於GoogleAI團隊發表的EfficientNet深度學習模型來進行自動光學檢查。進一步應用集成學習的方式將各模型預測的答案整理成單一預測結果。使得自動光學檢查準確率可以達到99.63%。 Automatic Optical Inspection methods based on deep learning have rapidly replaced traditional methods in recent years. How to improve the accuracy of defect detection is an important part of Automatic Optical Inspection. Ensemble learning is the application of multiple machine learning models to improve accuracy. How to integrate different deep learning models to improve accuracy is also an important topic for Automatic Optical Inspection. In this paper, I used transfer learning to train EfficientNet deep learning models published by the Google AI team for automatic optical inspection. Ensemble learning method is further applied to organize the answers predicted by each deep learning model into single prediction results. The accuracy of Automatic Optical Inspection defect inspection can reach 99.63%.
