摘要
2020年初開始，新冠肺炎(covid-19)疫情自中國武漢地區爆發後，便開始在全球迅速蔓延，在全球範圍內造成數以億計的人確診，及數以百萬計的人死亡，至今尚未平息，使得全球口罩需求快速增加，很多國家口罩需求大於口罩產能，導致口罩供不應求，價格飛漲，如何增加口罩產能平穩價格，成為世界各國必須解決的課題。 本論文提出三個方法，第一個方法結合開運算(Opening Operation)與ResNet152V2模型提出Image Preprocessing with ResNet152V2(IP-ResNet152V2)方法應用於口罩瑕疵檢驗系統，首先圖像預處理(Image Preprocessing, IP)的部分透過形態學中的開運算(Opening Operation)將圖像中的口罩定位並將口罩單獨從背景分離出來，接著透過ResNet152V2圖像分類網路進行口罩瑕疵檢測，根據我們的實驗結果其驗證集準確率為96.67%。第二個方法Image Preprocessing with InceptionV3 (IP-InceptionV3)延續第一個方法，進一步將圖像分類模型更換為InceptionV3，在參數量下降的情況下其驗證集準確率為97.56%。第三個方法Image Preprocessing with Xception(IP- Xception)延續我們提出的IP-InceptionV3方法，進一步將圖像分類模型更換為Xception，在參數量不變的情況下其驗證集準確率高於IP-InceptionV3方法來到了98.21%。 我們在實驗中也比較了多種模型有無加入圖像預處理(IP)其準確率與執行時間的變化，我們的實驗結果表明加入圖像預處理(IP)可將模型的辨識準確率提升且無明顯增加執行時間。
並列摘要
Since the beginning of 2020, the new crown pneumonia epidemic has spread rapidly around the world since it broke out in Wuhan, China. The global demand for masks has increased rapidly, and the demand for masks in many countries is greater than the production capacity of masks, resulting in a shortage of masks and soaring prices. To increase the production capacity of masks and stabilize the price has become a problem that all countries in the world must solve. This thesis proposes three methods. The first method combines the open operation and the ResNet152V2 model. We propose the Image Preprocessing with ResNet152V2(IP-ResNet152V2) method, which is applied to the mask defect inspection system. First, the image preprocessing part locates the mask in the image and it separates the mask from the background through the opening operation in the morphology. Furthermore, we use the ResNet152V2 image classification network to detect mask defects. Experimental results show that the validation set accuracy of IP-ResNet152 method is 96.67%. The second method is Image Preprocessing with InceptionV3(IP-InceptionV3) that it is based for the first method and further replaces the image classification model with InceptionV3. Experimental results show that validation set accuracy rate is 97.56% when the number of parameters decreases in IP-ResNet152 method. The third method is Image Preprocessing with Xception (IP-Xception) that it is based for the IP-InceptionV3 method, and it replaces the image classification model with Xception. The accuracy of the validation set is higher than that of the IP-InceptionV3 method when the parameters remain unchanged, reaching 98.21%. We also compared the changes in the accuracy and execution time of various models with or without image preprocessing. Experimental results show that adding image preprocessing of various models can improve the recognition accuracy without significantly increasing the execution time.