因著智慧城市與智慧家庭推動，人們越來越重視生活的品質，期待用科技改變我們的生活。而近年來隨著GPU進步與大數據時代的來臨，深度學習帶給各領域帶來革命性的進展，電腦視覺方面更是由深度學習引領著。科技始終來自人性，科技使我們的生活更便利。在我們生活週遭有著各式不同形態的門禁系統，從鑰匙、門禁卡、到生物特徵辨識。生物特徵辨識透過每個人獨有的特徵，不須另外攜帶任何形式的鑰匙。人臉由於不需接觸或是做出額外的動作，為所有生物特徵辨識中最為方便的方法，然而卻也是最為複雜的方法。 本論文提出透過深度網路，動態的對使用者進行人臉偵測與人臉辨識的門禁系統，不需特別停下進行辨識。人臉偵測採用卷積神經網路架構SSD(Single Shot MultiBox Detector)為主要方法，人臉辨識則採用VGGFace為主要方法。透過圖庫蒐集，資料增強，影像前後處理與實驗設計，訓練更為健全(robust)的人臉偵測與人臉辨識子系統。透過兩個系統的結合，便利用連續的影像，判斷是否為實驗室成員。連續的影像可避免單張影像誤測而造成的嚴重後果。本論文利用1280x960的彩色影片進行實驗測試，在GPU加速下可達到約30fps的速度。 ;Through the promotion of smart cities and smart families, people are paying more attention to the quality of life. They look forward to change the style of life by technology. The age of big data and GPU acceleration improvement brought deep learning a revolutionary progress in various fields, especially in computer vision. Technology derives from humanity, and technology makes our lives more convinent. There are various forms of access control system, such as keys, access cards, biometric identification, around us. Biometric identification distinguish different people by the unique characteristics of each person. Therefore users don’t need to bring any forms of keys anymore. Faces identification are the most convenient method for all biometric identification, because they don′t need to touch anything or make any extra moves. However, it’s the most complex method. We propose an access control system that performs face detection and face recognition dynamically, which makes users no need to stop for recognition. We use SSD(Single Shot MultiBox Detector)as the main model for face detection and VGGFace as the main model for face recognition. Through the collection of dataset, data augmentation, pre-processing of image, post-processing of image, and experimental design, we train the robust face detection and face recognition sybsystems. The system use continuous images as input to determine whether it’s a laboratory member. Using continuous frame as input can avoid false positive case make the system output wrong result. We use 1280x960 color video for experimental testing, and achieve about 30fps speed under GPU acceleration.
