使用基於秀拉顏色組成的多類藍雜訊取樣生成點描派畫作

Abstract

本論文提出了一個新的點畫技術，利用一個簡單且直觀的概念，可以有效地將輸入的彩色影像轉換成非擬真著色的彩色點畫輸出。首先，本論文藉由參考多幅秀拉的點畫作品，收集、分析、並模仿這些色點的顏色組成結構，進而創造出更多真實畫作中所缺乏的顏色，並將這些資訊儲存起來，以構成我們的色彩統計模型。接著，採用我們所修改的「多類藍雜訊取樣」作為色點的分佈方式，並參考我們在前處理時所建構的色彩統計模型，模仿秀拉的色點分布情況，而藍雜訊的性質確保色點之間的均勻以及隨機分布。 最後在實驗部分，我們使用多維度適合度檢定來分析我們與其他先前研究的輸出結果，比較每一個分割區域的色彩分佈與秀拉色彩分佈的相似度，證實了我們在顏色組成上與秀拉的作畫習慣最為近似。

In this thesis, we propose a new stippling technique, using a simple and intuitive concept to convert a color image into a pointillism painting. First, we collect, analyze, and imitate the color composition structure from Seurat‘s paintings. We further infer more color compositions, which do not contain in the reference painting, and include them in our color statistical model. Then, we use the modified multi-class blue noise sampling to distribute color points by looking up the color statistical model to imitate Seurat’s color composition. The blue noise property ensures that the color points are randomly located but remain spatially uniform. In our experiments, we use the multivariate goodness-of-fit tests to analyze our and other previous research’s results, comparing the color composition of each segmentation region to Seurat’s, and confirming that the color compositions of our results are most similar to Seurat’s painting habit.