預充式注射器是近代醫學中廣泛使用的醫療器械之一，為了提升醫療品質並獲得更健康、舒適的醫療服務，注射器結合人因工程領域的創新，發展出許多類型的注射型態以符合不同患者的需求。如果注射器內部的針劑瓶法蘭面破損，導致患者在緊急情況時無法將藥劑順利地注入體內，將會危及患者的生命安全，造成無可挽回的嚴重後果。 本研究旨在探討如何運用非接觸式的影像處理技術，檢測預充式針劑瓶的法蘭輪廓損傷，以AlexNet預訓練卷積神經網路進行遷移學習，將訓練資料分成完整輪廓、破損輪廓、無藥瓶三種類別，快速對影像的數據分類並且分析模型的精確度、召回率、F1-Score以及準確度，同時建立一套自動化檢測系統分析法蘭的損傷特徵，避免不符合規範的藥瓶流入市場。 研究結果顯示遷移學習應用於法蘭檢測具可行性，其正面法蘭模型的辨識準確度可達94.07%，背面法蘭模型的辨識準確度可達90.55%。經過法蘭輪廓檢測系統對遷移學習修正後，可將正面法蘭誤判成NOK的樣本總數從原先的3.7%降至2.2%，背面法蘭誤判成NOK的樣本總數從原先的6.5%降至2.6%，且實際為破損法蘭卻誤判成完整輪廓的影像皆能被檢測系統修正成正確的分類項目，最終正面法蘭影像得到97.77%的準確度，背面法蘭影像得到97.4%的準確度。;Nowadays, Prefilled syringes are widely used medical devices in modern medical community. In order to allow each patient to experience a comfortable medical quality, many different types of prefilled syringes have been developed in past decades. However, when a flange of the syringe is broken, the patient is not able to use the syringe in an emergency. This problem may affect the life with irreversible outcome and should be avoided. This research explores the methods using machine vision to detect flange edge in prefilled syringe. The AlexNet convolutional neural network is used for transfer learning and divides the data into three categories: normal flange image, Broken Flange image and no medicine bottle image. Confusion Matrix is the evaluation method of this research. The performance of the model is evaluated by four metrics: Precision, Recall, F1-Score, and Accuracy. The results of the research found that applying transfer learning to detect flanges is a feasible solution. The model accuracy of Front-Flange and Back-Flange can reach 94.07% and 90.55%. This research also developed a detection system, which can correct the damaged flanges misjudged as OK image. Finally, Front-Flange model got 97.77% accuracy and Back-Flange model got 97.4% accuracy.
