- 學位論文
基於深度學習之生成對抗網路技術以實現影像資料擴增
GAN based deep learning mechanism for image augmentation
摘要
影像辨識的發展趨勢逐漸普及，使得市場上各種領域的辨識技術扮演了重要的角色，藉由智慧化的影像辨識協助決策判斷，同時拓展更多元的應用。在訓練影像辨識模型時，通常會遇到訓練樣本不足或是類別的樣本不平衡問題，進而造成影像辨識的準確率不佳，通常問題發生的主要原因為某種類別的訓練資料不足，當我們的訓練資料集太少時，模型會變得無法順利地找出特徵，在沒有足夠特徵的情況下，進行分類或辨識，便導致辨識錯誤的問題。因此，為了解決此問題，本論文研究目的以車牌辨識為例，進行影像資料的擴增，產生更多的資料，彌補訓練資料不足的問題。於是本研究方法將提出利用少量的訓練資料，加入多模生成對抗網路技術，進而生成不同種類的訓練資料，並且保留影像的重要特徵。最後，本論文希望透過所提出的資料擴增的方法，能夠提升影像辨識的精確度。
並列摘要
The development of image augmentation is becoming more and more popular, making the recognition technique in various fields playing an important role in the market, assisting decision-making and judgment through intelligent image recognition and at the same time expanding more diverse applications. When training an image recognition model, it always encounters the problem including the lack of training data or the unbalance training data among different classes, which leads to the results of low accuracy in terms of image recognition. When the training data set is not enough or unbalanced, the model will unable to extract the features from the training data, leading to the difficulty of distinguishing the differences among various classes. To solve this problem, this research aims to use license plate recognition as an example to augment image data and make up for the lack of training data. In consequence, we proposed a small amount of training data and adopted the multimodal Generative Adversarial Networks technique to generate different kinds of training data and retain the important features of the image. Finally, it’s expected that the accuracy of image recognition can be improved by the proposed method of data augmentation.