在2019年底 新型冠狀病毒肺炎 (COVID-19)疫情爆發，全球口罩的需求瞬間大增，口罩的生產品質也變得重要。 因此本研究藉由 AI深度學習模型替口罩進行瑕疵檢測分類， 以當前精準 的卷積神經網路模型 EfficientNet 進行口罩檢測； 在使用模型時發現，雖然此類模型的準確率 高 ，但與 其他模型 相比則 需要較長的訓練時間 尤其是EfficientNet B7，在進行檢測 若沒有 高效能 顯示卡，就 無法進行訓練 ，最低也需要有一張 10G的顯示卡才能讓模型在不會將圖片縮得太小的情況下運作 。 於是本研究透過將EfficientNet B7的模型 神經層數量 減少， 並 觀察其在 此 情況下是否能夠 維持或提高檢測準確率 以目前修改表現最好的模型 得出 若 將 模型的 神經層數量減少百分之 60 不僅能降低顯示卡的負擔，還能減少訓練所要花費的時間， 且修改後的模型準確率與原始模型的準確率 也沒有太大差異 。 With the outbreak of the new coronavirus pneumonia (COVID-19) at the end of 2019, the global demand for masks has increased rapidly, and the production quality of masks has also become important. Therefore, in this study, the AI deep learning model is used to detect and classify the defects of masks, and the current accurate convolutional neural network model—EfficientNet is used for mask detection. EfficientNet has different models, from B0 to B7, and all models need to take a long time to train a prediction model with training data. Moreover, among these models, EfficientNet-B7 takes more times and gets better performance. However, if there is no high-performance video card, the prediction model cannot be trained. This is because when the images are shrank smaller, the performance of model will be worse. Hence, at least a 10G video card is required in this study. This study reduces the number of neural layers in the model of EfficientNet B7 and observes whether it can maintain or improve the detection accuracy in this way. The model with the best performance is modified and obtained. If the number of neural layers of the model is reduced by 60%, the results can be not only reduced the burden on the GPU, but also reduced the time spent on training model. And the accuracy of the modified model is not much different from that of the original model.
