|
|
|
 |
|
|
|
|
|
|
基於深度學習之敦煌壁畫復原之研究=Dunhuang Mural Restoration using Deep Learning |
|
|
|
| Author |
王瀚磊 (著)=Wang, Han-Lei (au.)
|
|---|
| Date | 2018 |
|---|
| Pages | 46 |
|---|
| Publisher | 國立臺灣大學 |
|---|
| Publisher Url |
https://www.ntu.edu.tw/
|
|---|
| Location | 臺北市, 臺灣 [Taipei shih, Taiwan] |
|---|
| Content type | 博碩士論文=Thesis and Dissertation |
|---|
| Language | 英文=English |
|---|
| Degree | master |
|---|
| Institution | 國立臺灣大學 |
|---|
| Department | 資訊工程學研究所 |
|---|
| Advisor | 李明穗、洪一平 |
|---|
| Publication year | 106 |
|---|
| Keyword | 敦煌石窟; 壁畫復原; 高解析度; 深度學習; 生成式對抗網路 |
|---|
| Abstract | 敦煌石窟藝術主要以建築、雕塑、壁畫三種藝術形式為主。隨著時間的流逝，現今的敦煌石窟藝術早已蒙受巨大的損壞，造成難以修復的情形。隨著科技的發展，許多人開始利用數位保留技術將現今的石窟保存，甚至可以透過虛擬實境技術觀察敦煌石窟之數位復原內容。透過將石窟建模、並將壁畫保存為紋理，我們可以在虛擬環境中觀察到石窟的現況，彷彿置身其中。然而我們仍然無法一窺當年石窟尚未損壞的盛景。
近代深度學習技術興起並被廣泛地利用在各種不同的領域，本篇論文係提出一套系統性的修復流程，針對高解析度之各類破損壁畫紋理進行修復。我們採用深度學習領域中之生成式對抗網路技術，廣泛讓電腦觀看損壞後的壁畫以及修復後的壁畫並且學習兩種影像之轉換。透過訓練此模型，我們可以同時產生高解析度但色彩不一致，以及低解析度但色彩一致之結果。利用影像處理技術即可結合高解析度以及顏色一致之特性，得到最佳之修復效果。
The art of Dunhuang Grottoes is mainly composed of three forms: building, sculpture, and murals. With the passage of time, the art of Dunhuang Grottoes has suffered from tremendous damage, and making it difficult to repair. With Virtual Reality, we can even observe the digital preservation of Dunhuang Caves in virtual environment as if we are in Dunhuang by modeling the caves and preserve mural as textures. However, we still cannot have a glimpse of how the grottoes look like without damage.
With the ever-changing nature of deep learning, it is widely used in many fields. In this work, we propose a systematic restoration process for high-resolution types of deteriorated mural textures. We use the Generative Adversarial Network technology in deep learning. By making the machine learn the transformation between deteriorated mural textures and restored mural textures, we can simultaneously produce high-resolution but color-inconsistent, as well as low-resolution but color-consistent results. After post-processing through digital image processing technology, we can get both high-resolution and color-consistent texture. |
|---|
| Table of contents | 誌謝 i
中文摘要 ii
ABSTRACT iii
CONTENTS iv
LIST OF FIGURES vi
Chapter 1 Introduction 1
Chapter 2 Related Work 4
2.1 Mural Restoration System 4
2.2 Digital Image Processing 4
2.3 Deep Learning 5
2.3.1 Auto Encoder 7
2.3.2 Generative Adversarial Network 8
Chapter 3 Dunhuang Mural Restoration 11
3.1 System Overview 11
3.2 Data Pre-processing 11
3.3 Training Process 15
3.3.1 Pre-train on Paired Training Data 16
3.3.2 Train on Whole Training Data 17
3.4 Inference Process 19
Chapter 4 Experiments 21
4.1 Observation 21
4.1.1 Padding 21
4.1.2 Resolution 22
4.1.3 Color Composition Preservation 24
4.1.4 Sharpness Enhancement 27
4.2 Classification First 30
Chapter 5 Conclusion and Future Work 32
5.1 Conclusion 32
5.2 Future Work 33
5.2.1 Super-resolution Generative Adversarial Network 33
5.2.2 Limited Pigment Constrain 35
Appendix 36
Comparison with L1 and L2 loss 36
61th Cave Dunhuang Mural Restoration 40
REFERENCE 44 |
|---|
| Hits | 250 |
|---|
| Created date | 2022.09.28 |
|---|
| Modified date | 2023.02.17 |
|---|
 |
Best viewed with Chrome, Firefox, Safari(Mac) but not supported IE
|
|
|
