生成對抗式網路(Generative Adversarial Network，GAN) 是目前人工智慧熱門的研究之一。GAN是一種強大的生成模型，其想法源自於博弈論的二人零和博弈，由一個生成器和一個判別器所組成，並透過對抗的方式來達到訓練的效果。在以CNN為主軸的影像辨識工作上，如何蒐集大量的影像以供網路的訓練及測試使用，是實務上常遇到的困難。在虹膜辨識的演算法開發上也是如此。如何能夠蒐集到足夠多的特例影像來重新訓練神經網路，是一個很重要的議題。虹膜影像本身不易收集，需要特殊的光學設備，而特殊情況的虹膜影像，則更難以收集。在實務上，特殊的虹膜影像常常使得虹膜辨識系統的某個環節失效(例如虹膜定位)。因此收集大量特殊的虹膜影像，對於訓練新的深度學習虹膜辨識演算法，會有極大的幫助。為了產生足夠真實且可以針對特定目的的影像，我們提出了一種新的生成對抗式網路，稱為 WGAN-GPAC (Wasserstein Generative Adversarial Networks with Gradient Penalty and Auxiliary Classifier)，在此新的GAN架構裡面，我們結合了WGAN-GP與一個獨立的分類器。運用此研究方法，根據自行設定的條件，WGAN-GPAC可以根據實驗者的需求，隨機的產生各種各樣特殊影像，藉此解決在進行深度學習實驗時訓練資料影像不足的問題。此方法不只能運用在虹膜辨識，也可以用於各種各樣以深度學習為主軸的影像資料庫的產生與建立。 ;Generative Adversarial Network (GAN) is one of the hot topics in artificial intelligence. GAN is a powerful generation model. The idea is derived from the two-person zero-sum game of game theory. It consists of a generator and a discriminator, which can be co-trained through the mutual competition.When using Deep Convolutional Neural Network (DCNN) as classifier, how to collect a large number of images for network training and testing is a common difficulty encountered in practice. Same difficulties happened for the development of deep learning algorithms for iris recognition. How to collect enough special images to retrain the neural network is an important issue. Iris images are not easy to collect and require special optical equipment, while iris images in special cases are more difficult to collect. In practice, special iris images often invalidate a certain part of the iris recognition system (e.g., iris localization). Therefore, collecting a large number of special iris images will greatly help to develop a deep-learning based iris recognition algorithm.In order to produce a sufficiently realistic and purpose-specific image, we propose a new architecture of GAN called WGAN-GPAC (Wasserstein Generative Adversarial Networks with Gradient Penalty and Auxiliary Classifier). In this new GAN architecture, we combined WGAN-GP with a auxiliary classifier. Using the proposed architecture, a variety of special images can be randomly generated according to the needs of the experimenter, thereby solving the problem of insufficient training image data during the deep learning experiment. This method can be used not only in iris recognition, but also in the generation and establishment of various image databases which is suitable for deep learning algorithm development.
