請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71361
標題: | 使用原始對偶演算法與顯著性篩選之灰階影像著色演算法 Example-Based Image Colorization Using Primal-Dual Algorithm and Saliency Screening |
作者: | 楊茜雯 Chien-Wen Yang |
指導教授: | 盧奕璋 Yi-Chang Lu |
關鍵字: | 灰階影像上色,一階原始對偶演算法, example-based image colorization,first-order primal-dual algorithm, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 影像著色演算法的目標是對一張灰階影像添加適當的色彩。傳統人工上色需要花費大量時間與人力,電腦輔助演算法的出現使得整個上色流程更有效率。由於其廣泛的應用,例如老舊照片修復、感測器融合等,近年來逐漸受到研究重視。然而影像上色本身是一個極度不適定且模稜兩可的問題,因為對於一張灰階影像,可能的色彩組合有非常多種。因此,使用者互動或是對於目標影像的先驗知識扮演了重要的角色,主要可分為三大類:基於塗色、基於參考圖像,以及全自動之上色方法。
本篇論文提出基於參考影像的著色演算法,給定一張內容相似的彩色參考圖像,可將其色彩轉移至目標影像,輸出視覺上自然合理的上色結果。本篇論文提出之方法結合了像素、超像素,以及基於圖像分割方法之優點。本篇論文有兩大特色:第一,不同於其他著色演算法中常見的特徵匹配流程,本論文提出了顯著性篩選流程,利用高層次的顯著性資訊來縮小匹配範圍,提升匹配準確度。第二,提出一個非凸的最佳化模型以輸出最終色彩。此模型包含三種邊緣感知的規律化項,可以在傳遞色彩的同時,保持亮度與色度通道間結構的一致性,因此可以有效解決區塊效應與光暈效應,並有很好的邊緣保持效果。最後,實驗結果顯示,與其他文獻中表現較佳的演算法相比,本文提出之演算法可以得到良好的上色結果。 Image colorization aims to add suitable chrominance values to a grayscale image. Since the traditional hand-coloring methods are time-consuming and require tremendous human efforts, computer-assisted algorithms have been invented to make the whole process more efficient. It has received increasing research attentions due to a wide range of applications, such as the restoration of old media and sensor fusion. However, the problem is inherently ill-posed and ambiguous since there are potentially many colors that can be assigned to a grayscale image. Hence, human interaction or prior knowledge of the target image is highly involved in the colorization process, and it is a challenging research topic worth further investigation. Previous methods fall into three main categories: scribble-based, example-based, and fully automatic colorization. In this paper, an example-based image colorization algorithm is proposed. Given a semantically-related color reference image, our method transfers colors from the reference to the target image, providing a visually plausible colorized result. We propose a colorization process leveraging the advantages of methods based on segmented regions, superpixel-level, and pixel-level. The algorithm has mainly two features. First, different from the typical feature matching process, we propose saliency screening, utilizing the high-level saliency information to narrow down the search space. This significantly improves the matching accuracy. Secondly, we propose a nonconvex variational model with edge-aware priors to choose the final color candidates. The proposed combination of priors propagates the colors while preserving mutual structures between luminance and chrominances. Our approach effectively suppresses blocky effects and halo effects, thus having good edge-preserving properties. Finally, experimental results demonstrate that the proposed algorithm outperforms the state-of-the-art methods. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71361 |
DOI: | 10.6342/NTU202004373 |
全文授權: | 未授權 |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-109-2.pdf 目前未授權公開取用 | 23.65 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。