基於深度學習工業異常檢測方法之研究

Abstract

隨著深度學習的技術日益成熟，將此技術應用於影像檢測的案例已經成為近年來學界與業界重要的研究議題。然而，過去的卷積神經網路檢測方法需要蒐集大量的瑕疵樣本圖像並消耗大量時間進行人工標記，這導致了深度學習導入實際產線的困難。有鑑於此，本研究提出了三個新的異常偵測模型，並探討其應用於實際產線檢測的效果。研究結果顯示所提出的模型在公開資料集 MVTec AD中比起過去的異常偵測方法有著更好的異常偵測能力。此外，在本研究中蒐集的兩組實際產線資料集中，由於其中的圖像更為複雜，因此所提出的方法優勢也更為明顯。更甚，本研究也探討了不同特徵抽取器對檢測效果的影響，結果顯示即使只使用較為輕量化的預訓練特徵抽取器，所提出的方法仍然具有良好的檢測能力，這代表了即使在算力受限的實際產線硬體中，所提出的方法仍然可以維持相當的檢測能力。最後，本研究也設計了應用程式軟體，讓使用者可以很簡單地訓練本研究所提出的模型，並與相機或網域內的硬體串接。這將使本研究可以更好地被應用於實際產線，真正為業界創造價值!

With the growing of deep learning techniques, applying this technology to image detection has become an important research topic in both academia and industry in recent years. However, previous CNN (Convolutional Neural Network) detection methods required collecting a large number of defective sample images and spending a lot of time on manual labeling, which made it difficult to implement deep learning in actual production lines. Therefore, this study proposes three new anomaly detection models and explores their effectiveness in real production line detection. The results show that the proposed models have better anomaly detection capabilities compared to previous methods on the publicly available dataset MVTec AD. Furthermore, in the two sets of actual production line datasets collected in this study, the advantages of the proposed methods are even more apparent due to the more complex images in these datasets. Moreover, this study investigates the impact of different feature extractors on detection performance. The results show that even with lightweight (require fewer computational resources) pre-trained feature extractors, the proposed methods still maintain good detection capabilities. This means that even in actual production line hardware with limited computing power, the proposed methods can still maintain considerable detection capabilities. Finally, an application software that allows users to easily train the proposed models and interface with cameras or hardware within the same internet domain was designed in this study. This will enable the application of the proposed methods in actual production lines, creating real value for the industry.