透過您的圖書館登入
IP:44.200.141.122
  • 學位論文

以深度學習為基礎之影像識別技術於教學管理之應用

Use of Deep Learning Based Face Recognition Approach in Teaching Management

指導教授 : 陳大正
若您是本文的作者,可授權文章由華藝線上圖書館中協助推廣。

摘要


本論文主要探討如何使用深度學習神經網路中的卷積神經網路透過即時攝影的方式持續性的觀察學生的學習狀態提供教師進行教學管理。 以往傳統的人臉辨識是透過幾何的方式去鎖定人臉五官的位置,容易受到遮蔽而造成辨識不良。除了需要正面面對攝影機,還必須清楚呈現五官的位置。若要實際應用在課堂上,學生的臉部可能因為不同的方向或狀態而由所遮蔽,無法準確地進行持續性的辨識。因此本研究使用一種透過擷取影像特徵用於圖像辨識的卷積神經網路,進行學生人臉的辨識。 但由於一間教室有多位學生,透過單一攝影機的拍攝,難以辨識所有的學生,且容易把背景的特徵也辨識進去而造成誤判。因此以往都是使用每個學生座位前放置一台攝影機,對攝影機中的人臉進行辨識。而本研究使用Faster R-CNN基於其區域建議網路對畫面中的人臉影像進行截取,再透過人臉辨識模型進行訓練。除了可以讓一張畫面中的多位學生進行身分辨識,更能夠排除掉不必要的背景,避免誤判。本研究所訓練之人臉身分辨識模型經測試後準確率高達99%,測試損失值為0.0278137。而除了畫面中的學生身分辨識外,本研究更透過情緒辨識模型去分析其上課時的當下情緒並記錄統計,去了解學生上課時的心情。 最終透過WebSocket傳遞參數驅動辨識,並以FLUX架構建立一套線上教學管理系統。將結果透過網頁的形式呈現出來,供教師進行教學管理評量之用途。

並列摘要


This essay explores the problem of fairness and stability that exist on the present roll call system. According to the past literature, most of the roll call systems were developed by using RFID technology and geometric face recognition technique. As we know several disadvantages of current way of roll call should be solved. For example, the traditional roll call in the classroom is time consuming; the problem of RFID roll call system may give someone the chance to swipe other person's card; and face recognition system cannot recognize the partially obscured image. It seems that very few literatures investigated the use of deep learning facial recognition in the teaching management. Therefore, we proposed that using deep learning based approach to recognize the students’ facial images so that attendance status of students can be then identified in classroom. Moreover, their emotions can be captured and analyzed by the proposed approach. We use RPN to improve the existing deep learning model to achieve better recognition accuracy. We proposed a new rule to improve the traditional roll call in classroom. It is based on the time that the students arrive in classroom to determine attendance rate and the time to stay in classroom, so that a fair and stable platform can be made. In this study, the modified convolutional neural network has been applied to increase the accuracy of facial recognition with/without partially obscured image. This study establishes a set of learning status judgment mechanism according to the ratio of students’ attendance time occupied by the total time, which is fairer than the traditional roll call in the classroom. In this thesis, a Teaching Management System (TMS) is proposed. By using the deep learning face recognition approach, the student's learning status can be revealed so that teachers can be with more supporting in their teaching in terms of classroom management.

參考文獻


史美瑤(2012)。21世紀的教學:以「學生學習為中心」的教師發展文。
林伯強(2016)。教育自主學習的利與弊。
陳舜芬,曾正宜,陳素燕 (2006)。我國大學生知識教育成果及其相關因素之研究。新竹教育大學學報NSC 94-2413-H-007-002,23,21-42。
Amari, S. (1998). Natural Gradient Works Efficiently in Learning. Neural Computation, 10(2), 251-276. http://dx.doi.org/10.1162/089976698300017746
Brunelli, R., & Poggio, T. (1993). Face recognition: features versus templates. IEEE Transactions On Pattern Analysis And Machine Intelligence, 15(10), 1042-1052. http://dx.doi.org/10.1109/34.254061

延伸閱讀