本論文用深度卷積生成對抗網路來產生假的圖片資料，並以此增加訓練分類模型時的訓練資料量。生成對抗網路藉由賽局理論的方式，讓鑑別網路與生成網路進行對抗訓練，由此來學習並模擬真實資料的分布。接著利用學到的分布和權重來將隨機亂數向量，映射為各類別的假圖片。本論文使用CIFAR-10 與MNIST 圖像資料庫。這兩個資料庫被廣泛應用於圖像辨識的研究，而本論文亦將使用其訓練資料對生成對抗網路和分類器做訓練，並以測試資料來做評估。首先以CIFAR-10 與MNIST的訓練集分別訓練一個架構為全卷積神經網路的分類器，並得到baseline分別為90.09%以
及99.61%。接著本論文用全卷積神經網路架構做為鑑別網路與生成網路，利用全卷積神經網路能有效學習圖像區域特徵的特性，讓鑑別網路能有更高的信心水準區分真圖片與假圖片的差異。同時生成網路為了騙過鑑別網路也會更精準的去模仿真圖片的細部特徵。本研究將嘗試以不同方法對圖像進行前處理，並用於訓練生成對抗網路。亦將比較各種圖像處理方式所訓練的生成對抗網路，其所產生假圖彼此間的差異性。接著使用訓練完成的生成網路產生假圖，以分類器及最近鄰演算法對假圖做挑選和標記。最後使用假圖來增加訓練資料量，以此幫助訓練分類模型。經由實驗結果證實，利用生成網路產生假圖增加訓練資料量此一方法，能有效提升辨識率。

In this paper, we use artificial examples synthesized by deep convolutional generative adversarial netwoeks in image recognition. First, class dependent generative adversarial networks (GAN) are trained through a game-theoretic setting for a game between Discriminator and Generator to learn the distributions of the classes, and then mapped the random vector to the artificial examples by the distributions of the classes. We use Use CIFAR-10 and MNIST Image database. These two databases are widely used in image recognition research. We also use its training data to train generative adversarial netwoeks and classifier, and use the test data to do assessment. We use CIFAR-10 and MNIST to trained an all convolutional network and get the baseline respectively 90.09% and 99.61%. Then, we used the all convolutional network architecture as Discriminator and Generator.The advantages of using the convolutional network, which learn the regional feature well, make Discriminator distinguish the difference between real data and artificial examples more accurately. Similarly, Generator will be more accurate to imitate the details of the real data in order to fool Discriminator. We try to preprocess the image in different ways and use it for training generative adversarial netwoeks. Then use the trained network to generate artificial examples, and use the classifier and the nearest neighbor algorithm to select the artificial examples. Finally, we use artificial examples to increase the amount of training data and train the classifier. The experimental results show that this method can effectively improve the recognition rate.

論文審定書. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .i
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ii
摘要. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .iii
ABSTRACT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Table of Contents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Tables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Figures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
Chapter 1 簡介. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1
1.1 背景與研究動機. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 文獻回顧. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 MNIST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 CIFAR-10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.3 生成對抗網路(Generative Adversarial Networks, GAN) . . . . . . . 4
Chapter 2 基本架構與工具介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6
2.1 生成對抗網路(Generative Adversarial Networks, GAN) . . . . . . . . . . . 6
2.1.1 Discriminator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 最近鄰演算法(K nearest neighbor, KNN) . . . . . . . . . . . . . . . . . . . 7
2.3 卷積神經網路（Convolutional Neural Network, CNN） . . . . . . . . . . . 8
2.4 工具介紹:Tensorflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Chapter 3 資料庫介紹與研究方法. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10
3.1 資料庫介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.1 MNIST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.2 CIFAR-10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 全卷積神經網路(All Convolutional Net) . . . . . . . . . . . . . . . . . . . 11
3.3 批量正規化(Batch Normalization) . . . . . . . . . . . . . . . . . . . . . . . 12
3.4 深度卷積生成對抗網路(Deep Convolutional Generative Adversarial Netwoeks)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Chapter 4 實驗. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17
4.1 實驗設定. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 基準實驗. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2.1 CIFAR-10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2.2 MNIST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 生成樣本實驗. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3.1 原圖訓練DCGAN實驗. . . . . . . . . . . . . . . . . . . . . . . . 22
4.3.2 先影像處理再訓練DCGAN . . . . . . . . . . . . . . . . . . . . . . 24
4.3.3 先rescale再訓練DCGAN . . . . . . . . . . . . . . . . . . . . . . . 26
4.3.4 影像處理加rescale再訓練DCGAN . . . . . . . . . . . . . . . . . . 27
4.3.5 MNIST訓練DCGAN實驗. . . . . . . . . . . . . . . . . . . . . . . 27
4.3.6 評估DCGAN訓練收斂情形. . . . . . . . . . . . . . . . . . . . . . 29
4.4 分類模型評估生成樣本. . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.5 最近鄰法評估生成樣本. . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.6 生成樣本增加訓練資料. . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.7 挑選生成樣本再加入訓練集實驗. . . . . . . . . . . . . . . . . . . . . . . 36
4.7.1 分類模型挑選生成樣本. . . . . . . . . . . . . . . . . . . . . . . . 36
4.7.2 最近鄰法挑選生成樣本. . . . . . . . . . . . . . . . . . . . . . . . 38
4.8 加入水平翻轉訓練資料與生成樣本實驗. . . . . . . . . . . . . . . . . . . 39
4.8.1 訓練資料水平翻轉實驗. . . . . . . . . . . . . . . . . . . . . . . . 40
4.8.2 水平翻轉資料加分類模型挑選生成樣本實驗. . . . . . . . . . . . 40
4.8.3 水平翻轉資料加最近鄰法挑選生成樣本實驗. . . . . . . . . . . . 41
4.8.4 實驗結果統整. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.9 MNIST實驗. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Chapter 5 結論與未來展望. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .43

[1] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to
document recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998.
[2] D. Ciregan, U. Meier, and J. Schmidhuber, “Multi-column deep neural networks for
image classification,” in Computer Vision and Pattern Recognition (CVPR), 2012 IEEE
Conference on, pp. 3642–3649, IEEE, 2012.
[3] L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus, “Regularization of neural networks
using dropconnect,” in Proceedings of the 30th International Conference on Machine
Learning (ICML-13), pp. 1058–1066, 2013.
[4] D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber, “Deep big simple
neural nets excel on handwritten digit recognition, 2010,” Cited on, vol. 80.
[5] B. Paszkowski, W. Bieniecki, and S. Grabowski, “Preprocessing for real-time handwritten
character recognition,” in Computer Recognition Systems 2, pp. 470–476, Springer,
2007.
[6] T.-H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “Pcanet: A simple deep learning
baseline for image classification?,” IEEE Transactions on Image Processing, vol. 24,
no. 12, pp. 5017–5032, 2015.
[7] J.-R. Chang and Y.-S. Chen, “Batch-normalized maxout network in network,” arXiv
preprint arXiv:1511.02583, 2015.
[8] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller, “Striving for simplicity:
The all convolutional net,” arXiv preprint arXiv:1412.6806, 2014.
[9] I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Bengio, “Maxout
networks,” arXiv preprint arXiv:1302.4389, 2013.
[10] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional
neural networks,” in Advances in neural information processing systems,
pp. 1097–1105, 2012.
[11] S. Zagoruyko, “92.45% on cifar-10 in torch,” Torch Blog, 2015.
[12] B. Graham, “Fractional max-pooling,” arXiv preprint arXiv:1412.6071, 2014.
[13] C. Szegedy,W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus,
“Intriguing properties of neural networks,” arXiv preprint arXiv:1312.6199, 2013.
[14] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural networks are easily fooled: High
confidence predictions for unrecognizable images,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 427–436, 2015.
[15] P. Tabacof and E. Valle, “Exploring the space of adversarial images,” in Neural Networks
(IJCNN), 2016 International Joint Conference on, pp. 426–433, IEEE, 2016.
[16] I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial
examples,” arXiv preprint arXiv:1412.6572, 2014.
[17] T. Miyato, A. M. Dai, and I. Goodfellow, “Adversarial training methods for semisupervised
text classification,” arXiv preprint arXiv:1605.07725, 2016.
[18] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,
A. Courville, and Y. Bengio, “Generative adversarial nets,” in Advances in neural information
processing systems, pp. 2672–2680, 2014.
[19] E. L. Denton, S. Chintala, R. Fergus, et al., “Deep generative image models using
alaplacian pyramid of adversarial networks,” in Advances in neural information processing
systems, pp. 1486–1494, 2015.
[20] J. Gauthier, “Conditional generative adversarial nets for convolutional face generation,”
Class Project for Stanford CS231N: Convolutional Neural Networks for Visual Recognition,
Winter semester, vol. 2014, p. 5, 2014.
[21] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep
convolutional generative adversarial networks,” arXiv preprint arXiv:1511.06434, 2015.
[22] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, and R. Webb, “Learning
from simulated and unsupervised images through adversarial training,” arXiv preprint
arXiv:1612.07828, 2016.
[23] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by
reducing internal covariate shift,” arXiv preprint arXiv:1502.03167, 2015.