利用深度學習辨別繪圖之風格及美學性質

Abstract

本篇論文主要在探討當給定一張佈局演算法不明的圖結構視覺化圖片，在缺乏其他繪製相同圖結構但不同佈局演算法的結果來一起比較時，為了判斷優劣，能以本篇所提之方法做圖形繪製風格判定，進而推估此張佈局演算法不明的圖片所具有之美學特性。圖形繪製是一項研究多時的領域，現有的圖形佈局演算法通常不能優化所需的所有美學特性，因此在對這些眾多演算法所繪製出的結果做評估時，除了一些既定的美學指標計算外，常做的就是所謂肉眼直接觀察的人工判讀。但通常都是在給定幾種圖形佈局方法後，從複數張給定圖片做優劣的選擇，如果只單單就一張圖形繪製結果來看的話便無從對其有任何的判讀。本論文中提出了一種基於深度學習的方法，將大量預挑選的圖形佈局演算法所繪製出的圖片用於卷積神經網路分類架構的訓練。根據訓練之後的分類器用於分類各式圖形佈局演算法或圖形佈局演算法不明的資料，經分類測試後所呈現之美學性質結果，可發現類似風格的圖形佈局演算法會有近似的美學性質表現，而佈局法不明的圖片經由被分類後所推估之美學性質會大致與直接計算的數值結果相符。總而言之，本論文所提之方法是利用相似的佈局風格會被神經網路歸類於一類的性質，並藉由觀察分類後結果各自算出之美學指標的數值便能得知各類演算法的特性，之後拿著佈局方法不明的圖片就能以神經網路分類並找到相似風格的圖形佈局演算法進而得出其所擁有的美學特性。

We propose a method to determine some aesthetic properties of an image which is drawn by using unknown graph layout algorithms. Graph drawing has been studied for decades. There is no layout algorithm which can satisfy all of the important aesthetic properties for graph layouts. Therefore, when evaluating any of these algorithms, user study is the way to determine the performance. When conducting user study, people need to choose the best layout image from the results generated by selected layout algorithms. It is hard for users to determine whether an image drawn by an unknown layout algorithm performs good or not if there is no other layout image to compare with it. In this thesis, we propose a deep learning based method which is a convolutional neural network classifier trained by images drawn in selected graph drawing algorithms. After testing the model by the dataset composed of images drawn by unknown layout algorithms, we find the graph layout algorithms of similar styles have similar aesthetic properties, and the estimated aesthetic properties of the images drawn by unknown layout methods will nearly be consistent with the directly calculated numerical results. In summary, the method proposed in this thesis is to use the tendency that similar layout style will be classified into the same class by the neural network, and the characteristics of various algorithms can be learned by observing the values of the aesthetic indicators of the classification results. Our study shows that, given an image drawn by unknown layout methods, we can use neural networks to classify it and find a similar style of a graph drawing algorithm to obtain its aesthetic properties.