基於卷積神經網路的門禁與自動標記系統

Abstract

隨著科技越來越進步，人工智慧也隨之崛起。生活周遭的事物越發自動化與智慧化，門禁系統做為住戶的安全管理員，在智慧化家庭中的地位舉足輕重。對於進出建築物或私人住所，人們為了追求快速、方便而漸漸摒棄傳統的鑰匙，改用磁釦或是門禁卡。甚至連生物特徵識別技術，如指紋辨識、虹膜辨識、人臉辨識也越來越普及化。人臉辨識是所有辨識法中最為方便的一種，卻也是最複雜的一種。其容易受外在環境、光線、人臉角度與表情等影響辨識結果，因此常常需要結合多種複雜方法進行人臉辨識，造成處理速度上較為緩慢，進而難以應用於需要及時辨識的門禁系統上。本論文採用卷積神經網路架構 YOLO("You Only Look Once")為主軸方法，藉由資料蒐集、影像的前後處理與設計實驗，訓練出人臉偵測、人臉辨識、人臉驗證等三個功能的模型。結合上述三種功能，並利用影片串流的特性，解決單張影像辨識率較差的缺點。達成具有即時辨識人臉是否為家庭成員，並給定非家庭成員專屬標籤功能的門禁系統。

With the progress technology, artificial intelligence is rising. The things around life become more automated and intelligent. Access control system is the most important part of smart home because it relates to the safety of the householders. People gradually abandon the keys, and use the key cards in pursuit of accessing to buildings or homes quickly. Even biometrics become more popular, such as fingerprint identification, iris recognition, and face recognition. Face recognition method is the most convenient of all methods, but also the most complex one. It is easily influenced by external environment, light, face angle and expression. So we often need to use a variety of complex methods for face recognition. That cause processing speed slowly and difficult to apply for access system which require to identity identification on real time. We use a convolutional neural network architecture "YOLO"(You Only Look Once) as the methodology of thesis. By data collecting, image processing and experimental designing, we train these models including face detection, recognition, and verification. We uses the above three models and the characteristics of video streaming to solve the shortcomings that single image recognition rate is poor. Our access control system can identify family members on real time and it can also automatically label non-family members.