深度學習是近幾年非常熱門的研究方向，也應用在許多的領域上，尤其在機器學習方面不管是在自動化產業上、工業4.0、語音辨識等都能見到蹤影，其中影像辨識方面在近年來各方面的應用上被廣泛使用，相關研究也如雨後春筍般被提出且趨於成熟。本實驗動機在於藉由學校宿舍出入門口仍須透過鑰匙或感應卡才可通行此原因，希望能設計出一個架設在門口的感測器，透過鏡頭感測通行者的臉部特徵來檢測出其性別或是身分最後判斷是否可通行。

本論文實驗方法在建立模型，透過深度學習其中之一的YOLO2類神經網路架構作為基底搭載在NVIDIA Jetson TX2前端開發模板上，實驗全程皆在此平台上執行，並透過自行收集的資料集以及修改卷積神經網路的參數設定，最後在NVIDIA Jetson TX2上自行訓練，建構出兩個可以辨識人臉的性別與身分的目標檢測模型，本篇論文主要講解如何建立模型、訓練集以及YOLO2網路架構的設定與訓練。

實驗結果在性別與身分兩個模型於自己建立的資料集訓練下辨識平均準確率可達到百分之90以上，並且在影格率約10到20的情況下，卷積神經網路的輸入層可達608×608甚至到832×832的圖像大小。

Deep learning has been a very popular research direction in recent years, and it can also applied to many fields, especially in machine learning, either in the automation industry, Industry 4.0, speech recognition, or other areas. The motivation for this research paper is that the entrance to school dormitories still requires a key or a proximity card to enter. I hope to design a sensor installed in the doorway to detect people passing by the lens and capturing facial features in order to detect people’s gender or identity and finally determine whether they can pass or not.

The experimental method of this thesis is to build a model and use one of the YOLO2 neural network architectures as the base to carry the NVIDIA Jetson TX2 as the base. The parameter settings of the convolutional neural network are modified. And trained on the NVIDIA Jetson TX2 to construct two target detection models that can detect gender and identity. This paper mainly explains how to build models and training sets by means of definitions and collections, as well as the configuration and training of the YOLO2 network architecture.

The experimental results show that the mean average precision (mAP) of the two models of gender and identity can reach 90 or more in real time while the image size of the input layer can reach 608 × 608 or even 832 × 832.

摘要 I
ABSTRACT II
致謝 III
目錄 IV
圖目錄 VI
表目錄 IX
第一章 緒論 1
1.1 前言 1
1.2 研究動機 2
1.3 論文架構 4
第二章 深度學習與卷積神經網路 5
2.1 前言 5
2.2 深度學習 5
2.3 卷積神經網路 6
2.4 R-CNN 10
2.5 Fast R-CNN 13
2.6 Faster R-CNN 15
2.7 YOLO 17
2.8 YOLO2 24
第三章 實驗設計與結構 35
3.1 前言 35
3.2 實驗環境 35
3.3 Darknet 37
3.4 實驗流程 39
3.5 資料集建立 39
3.5.1 性別資料集 40
3.5.2 身分資料集 41
3.6 資料標籤化 42
3.7 訓練集與測試集 44
3.8 建立設定檔資料夾 46
3.9 參數設定 50
3.10 訓練 51
第四章 實驗結果與分析 55
4.1 前言 55
4.2 mAP(mean Average Precision) 55
4.3 人臉性別辨識實驗結果 56
4.4 人臉身分辨識實驗結果 64
第五章 結論與未來展望 68
5.1 結論 68
5-2 未來展望 69
參考文獻 70

[1] Kobe Chen.(2016).泛科技3分鐘搞懂深度學習到底在深什麼。檢自https://panx.asia/archives/53209 (Jan.8,2020)
[2] 尹相志(2016)。關鍵評論AlphaGo到底是如何下棋？兼談幾個對AlphaGo的誤解.檢自https://www.thenewslens.com/article/38066(Jan.8,2020)
[3] Neil.(2017). Predicting Steering Angles with Deep Learning Part 2。檢自 https://neilnie.com/2018/01/27/predicting-steering-angle-with-deep-learning-part-2/ (Jan.8,2020)
[4] 維基百科深度學習。檢自https://zh.wikipedia.org/wiki/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0(Jan.4,2020)
[5] 周秉誼(2016)。淺談Deep Learning原理及應用。檢自http://www.cc.ntu.edu.tw/chinese/epaper/0038/20160920_3805.html(Jan.4,2020)
[6] Tommy Huang.(2018).卷積神經網路(Convolutional neural network, CNN) — CNN運算流程。檢自https://reurl.cc/XXnALR(Jan.4,2020)
[7] Brandon Rohrer.(2016).How do Convolutional Neural Networks work?檢自https://brohrer.github.io/how_convolutional_neural_networks_work.html(Nov.17,2019)
[8] K.E.A. van de Sande,”Selective Search for Object Recognition”,2011.
[9] Ross Girshick, Jeff Donahue, Trevor Darrell, &Jitendra Malik,”Rich feature hierarchies for accurate object detection and semantic segmentation”,2013
[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, &Jian Sun,”Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition”,2014
[11] Ross Girshick,”Fast R-CNN”,2015
[12] Shaoqing Ren, Kaiming He, Ross Girshick, &Jian Sun,”Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks”,2015
[13] Joseph Redmon, Santosh Divvala, Ross Girshick, &Ali Farhadi,”You Only Look Once: Unified, Real-Time Object Detection”,2015
[14] NMS—非極大值抑制演算法的理解。檢自https://www.itread01.com/content/1541722582.html(Jan.4,2020)
[15] Joseph Redmon, &Ali Farhadi,”YOLO9000: Better, Faster, Stronger”,2016
[16] 陳穗碧(2019)。[Meetup5_2019] YOLOv2 - 陳穗碧(Mora chen)。檢自https://www.youtube.com/watch?v=xo_iUBjz0F4&list=PLZmDTa70TNq2siZwwoeQkDaUBbBmBtNn1&index=7&t=0s.(Nov,7.2019)
[17] Joseph Redmon, &Ali Farhadi,”YOLO9000: Better, Faster, Stronger”,2016
[18] hrsstudy(2017) ,YOLOv1論文理解。檢自
https://www.pianshen.com/article/2428175083/(Nov,7.2019)
[19]測試影片：ASIAN | How You See Me.檢自
https://www.youtube.com/watch?v=OTLtjGUGgN8&list=PLZmDTa70TNq2siZwwoeQkDaUBbBmBtNn1&index=6&t=0s(Nov,1.2019)