台灣製藥產業對於藥錠的表面瑕疵檢測多以人力進行檢測，利用人工檢測的方式不但缺乏時間效率，且容易有不穩定及誤判之問題。近年來深度學習的發展快速，使得神經網路也逐漸應用在許多領域。如想基於卷積神經網路(Convolutional Neural Networks, CNN)為主軸架構來執行瑕疵檢測，則須提供足夠數量的缺陷樣本才能達到理想的訓練目標。惟缺陷充分樣本很難克服，且需花費大量時間對瑕疵做人工標記動作。
本論文研究使用生成對抗網路(Generative Adversarial Network, GAN)，僅使用正常藥錠表面影像的情況下來進行神經網路模型訓練，同時引進Wasserstein GAN(Wasserstein Generative Adversarial Networks, WGAN)和自編碼器(Autoencoder)來重新建構一個用於影像重建的生成對抗網路，藉由比較影像重建的前後來達成瑕疵檢測。由於生成對抗網路對於較小面積瑕疵的誤判率較高，因此本研究針對黑點瑕疵，也提出利用傳統光學檢測進行檢測的演算法。一系列的實驗證實本研究提出的演算法能夠發揮很高的瑕疵檢出率。

In Taiwan, pharmaceutical industries generally inspect surface of tablets for defects manually. This will result in not only time-consuming but also undesirable misjudgments. In recent years, due to the fast development of deep learning, Neural Network has been applied to more and more fields. In order to train the Convolutional Neural Networks for the usage of defects detection, a large number of defective samples have to be provided. However, it is very difficult to collect enough defective samples, and it also takes enormous amount of time to mark the defects manually.
This research makes use of Generative Adversarial Network(GAN) to train the neural network model by only providing images of normal tablets. At the same time, Wasserstein Generative Adversarial Network(WGAN) and Autoencoder are used to rebuild a GAN for image reconstruction, comparing the image before and after reconstruction to detect the defects. Because of GAN fails to detect small defect area, this research also implements traditional optical inspection techniques to inspect the defect of black spots. A series of experiments proves that the algorithms developed in this thesis is able to give high defect inspection rate.

[1]I. Goodfellow et al., "Generative adversarial nets," in Advances in neural information processing systems, 2014, pp. 2672-2680.
[2]M. Arjovsky, S. Chintala, and L. J. a. p. a. Bottou, "Wasserstein gan," 2017.
[3]M. Arjovsky, S. Chintala, and L. Bottou, "Wasserstein generative adversarial networks," in International conference on machine learning, 2017, pp. 214-223.
[4]I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Courville, "Improved training of wasserstein gans," in Advances in neural information processing systems, 2017, pp. 5767-5777.
[5]P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, "Image-to-image translation with conditional adversarial networks," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1125-1134.
[6]Z. Zhao, B. Li, R. Dong, and P. Zhao, "A Surface Defect Detection Method Based on Positive Samples," in Pacific Rim International Conference on Artificial Intelligence, 2018, pp. 473-481: Springer.
[7]D. C. Plaut, "Experiments on Learning by Back Propagation," 1986.
[8]Y. Freund and R. E. J. M. l. Schapire, "Large margin classification using the perceptron algorithm," vol. 37, no. 3, pp. 277-296, 1999.
[9]D. P. Kingma and J. J. a. p. a. Ba, "Adam: A method for stochastic optimization," 2014.
[10]I. Sutskever, J. Martens, G. Dahl, and G. Hinton, "On the importance of initialization and momentum in deep learning," in International conference on machine learning, 2013, pp. 1139-1147.
[11]J. Duchi, E. Hazan, and Y. J. J. o. M. L. R. Singer, "Adaptive subgradient methods for online learning and stochastic optimization," vol. 12, no. Jul, pp. 2121-2159, 2011.
[12]Y. LeCun, L. Bottou, Y. Bengio, and P. J. P. o. t. I. Haffner, "Gradient-based learning applied to document recognition," vol. 86, no. 11, pp. 2278-2324, 1998.
[13]S. Ioffe and C. J. a. p. a. Szegedy, "Batch normalization: Accelerating deep network training by reducing internal covariate shift," 2015.
[14]C. Gonzalez and E. Woods, "Digital Image Processing, Addison-Wesley Publishing Company," 1991.
[15]機. 陳賢義 自動化感測技術專輯, "視覺技術在線上檢測之應用," vol. 6, pp. 267-279.
[16]J. Long, E. Shelhamer, and T. Darrell, "Fully convolutional networks for semantic segmentation," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 3431-3440.
[17]M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, "Deconvolutional networks," in Cvpr, 2010, vol. 10, p. 7.
[18]J.-Y. Wu, "以生成對抗式網路產生特定目的影像—以虹膜影像為例," National Central University, 2018.
[19]李孟坤 and 彭德保, "高分子有機發光二極體瑕疵自動光學檢測系統," 2004.
[20]林義祥 and 彭德保, "機器視覺應用於照相手機鏡片模組邊緣瑕疵檢測," 2011.
[21]陳明翰, "基於生成對抗網路之圖像彩色化," 2018.