| 摘要: | 人工智慧是電腦科學領域近年廣泛討論的議題之一，透過的對訓練資料擷取特徵學習，使用機器學習自主創作的技術也蓬勃發展，其中發展最佳的領域莫過於圖像生成。
本篇論文著重於自動創作生成卡通圖像，已將同種類的訓練資料透過分割、分群並取得區域部位之間的鄰近相連關係(Region Relationship Graph)為前提，後續將各個區域部位(Region)實作修補、形變、組合等等，創作出全新的圖像，相較於生成對抗網路(Generative Adversarial Network，GAN)使用深度學習的神經網路，本篇採用影像處理的方法實作，較能在運算時間、資料集數量與硬體資源上取得優勢。
本篇論文所提出的系統分為三個階段，考慮到原始輸入圖像之中Region與Region有遮蔽的影響，分割出來的Region會有被遮蔽的凹陷處，系統第一階段是將各個Region做修補的動作。為了創作的多樣性，第二階段是將各Region做形變的動作。最後則是隨機選擇一種模板，將形變後的Region組裝並調整。其中使用者也能調整參數，其系統自帶的隨機參數與使用者的參數會交叉出無數種組合，如此創作出全新的圖像。
在實驗中可見，其創作的訓練資料只要數張，就能有全新的創作，越多的訓練資料能有更多樣性的創作。而結果也能辨別出與訓練資料是同種類的物件。
;Artificial intelligence is one of the topics that have been widely discussed in the field of computer science in recent years. Through the acquisition of feature learning from training data and using machine learning to create, the technology has also flourished. Among them, the best field for development is image generation.
This paper focuses on the automatic creation of cartoon images. It has premised that the same kind of training data is divided, grouped, and acquired the region relationship graph. Subsequent implementation of each region. Patching, deformation, assembling, and so on, create new images. Compared to the neural network that uses deep learning in Generative Adversarial Network (GAN), this paper adopts image processing method to implement it. Take advantage of computing time, data sets, and hardware resources.
The system proposed in this paper is divided into three stages. Considering that Region and Region in the original input image have covering effects, the segmented region will have shadowed depressions. The first stage of the system is to inpaint each region. For the diversity of creation, the second stage is deforming the regions. Finally, a template is randomly selected, and the modified Region is assembled and adjusted. The user can also adjust the parameters. The random parameters of the system and the user′s parameters will cross the countless combinations, creating a new image.
It can be seen in the experiment that if using a few training data, they can have new creations, and more training materials can create more diversity. The results can also identify objects of the same kind as the training data. |
