基於深度學習物件偵測網路之口罩佩帶辨認

MASK WEARING IDENTIFICATION BASED ON DEEP LEARNING OBJECT DETECTION NETWORKS

佩帶好口罩有助於減少經由呼吸道飛沫傳染的風險，確保個人與他人的健康。研發口罩佩帶辨認系統有助於落實防疫自動化，降低人力成本，提高檢測效率，並降低檢測人員的感染風險。本研究採用最先進的通用深度學習影像物件偵測網路（像是YOLOv5）來自動辨認口罩的佩帶狀態，包括good（有正確戴好口罩）、improper（露出鼻頭，但未露出嘴巴）以及bad（沒有戴上口罩）。本研究使用現有的口罩遮擋人臉的資料庫訓練神經網路，再以手機實際錄製開放空間中的影片為測試集，以評估訓練好的神經網路對口罩佩帶辨認的效能。實驗結果顯示平均正確率為99.6%，平均精確度為99.7%，平均召回率為99.7%且平均執行速度為每秒66.3個畫面，整體表現已達可有效輔助人工檢測的水準。

Wearing a mask can help reduce the risk of droplets from the respiratory tract and ensure the health of individuals and others. The research and development of the mask wearing identification system will help implement the automation of epidemic prevention, reduce labor costs, improve detection efficiency, and reduce the risk of infection for inspectors. This study uses the most advanced general-purpose deep learning image object detection network (such as YOLOv5) to automatically identify the wearing states of the mask, including good (the mask is properly worn), improper (the nose is exposed, but the mouth is not), and bad (not wearing a mask). This study uses the existing database of mask-concealed human faces to train the neural network, and then uses a mobile phone to actually record videos in the open space as the test set to evaluate the effectiveness of the trained neural network for mask wearing recognition. Experimental results show that the average accuracy is 99.6%, the average precision is 99.7%, the average recall is 99.7%, and the average execution speed is 66.3 frames per second. The overall performance has reached a level that can effectively assist manual inspection.