中原大學電機工程學系
Loading
專題成果首頁>專題成果
107學年度 >訊號與系統組
Van Gogh is risen!以人工智慧深度學習進行畫風轉移
指導老師:洪穎怡   組長:顏暐庭   組員:莊晉翔、簡琪臻
鑒於現今攝相和修圖程式的發達,加上人工智慧技術日益白熾,本研究將深入探討畫風轉移的過程。人工智慧包含人工神經網路(Artificial Neural Network, ANN)、卷積神經網路(Convolutional Neural Networks, CNN)、遞歸神經網路(Recurrent/Recursive Neural Network, RNN)等神經網路。而應用於畫風轉移、最具代表性的神經網路為經典卷積神經網路模型-VggNets,這是一個在網路上已經訓練完成的深度神經網路模型,透過此模型,本研究將能進一步探討畫風轉移的內部程式之運作模式。
為了比較各個以訓練之神經網路模型的差異性,本研究使用三種神經網路-AlexNet、VGG16和VGG19來進行畫風轉移,利用卷積層抓取風格圖的特徵和內容圖的特徵結合後輸出成合成圖。合成圖中線條的細膩程度會因卷積層的數量而變化,實驗結果為卷積層使用的數量越多輸出結果越理想,因此以VGG19的16層卷積層輸出結果最佳。本研究在合成圖產生後再經影像處理還原內容圖的原本顏色,使最終輸出結果能基於風格圖的線條紋理產生內容圖的顏色與景物。
In view of the development of current photogrammetry and retouching programs, and the increasing incandescence of artificial intelligence technology, this study will delve into the process of artistic style transfer. Artificial intelligence includes neural networks such as Artificial Neural Network (ANN), Convolutional Neural Networks (CNN), and Recursive Neural Network (RNN). The most representative neural network used in the artistic style transfer is the classical convolutional neural network model-VggNets, which is a deep neural network model that has been trained on the network. Through this model, this study will be able to further explore the mode of operation of the internal program of the artistic style transfer.
In order to compare the differences between the various trained neural network models, the study uses three neural networks - AlexNet, VGG16 and VGG19 for the artistic style transfer, using the features of the convolution layer to capture the features of the style map and the features of the content map into a composite map. The degree of detail of the lines in the composite image varies depending on the number of convolution layers. The results show that using the more number of convolution layers, the output results are much better. Therefore, the output of using 16 convolution layers of VGG19 is the best. In this study, the original color of the content map is restored by image processing after the composite image is generated, so that the final output result can generate the color and scene of the content map based on the line texture of the style map.
我們決定將壓電陶瓷晶片和鞋子做結合。
   
 
32023 桃園市中壢區中北路200號 No-.200, Zhongbei Rd., Zhongli City, Taoyuan County 320, Taiwan(ROC) TEL:03-2654801 FAX:03-2654899