指導教授：丁建均臺灣大學：電信工程學研究所賴威昇Lai, Wei-ShengWei-ShengLai2014-11-302018-07-052014-11-302018-07-052014http://ntur.lib.ntu.edu.tw//handle/246246/264263影像去模糊是電腦視覺和影像處理領域中一個基本的問題。影像去模糊的目標是從輸入的模糊影像恢復模糊函數和潛在的原始影像。它是類似於影像去雜訊和影像超解析技術的逆問題，但因為運動模糊函數和自然影像的多樣性而更具挑戰性。近年來解影像去模糊的方法包含兩個主要階段：估計模糊函數和復原影像。模糊函數估計演算法必須依靠輸入影像有足夠的顯著邊緣來推斷一個可靠的模糊函數。大多數的方法僅使用YUV顏色空間中的Y層（亮度層）來估計模糊函數。然而，研究指出RGB到Y的轉換可能會喪失對比度和邊緣的資訊。 在本篇論文中，我們專注於解決由手持相機而產生晃動所拍攝的單張影像去模糊問題。我們提出一種新的去模糊演算法，在估計模糊函數的階段同時考慮RGB顏色通道的資訊。我們也提出了兩個從色彩線模型推導而來的彩色影像先驗機率(color-image priors)。我們的彩色影像先驗機率不僅能恢復影像邊緣的對比度，還可以降低最佳化過程中由不準確的模糊函數所產生之影像雜訊。實驗結果顯示我們的演算法可增強估計模糊函數的穩定性和準確性，精確的模糊函數可以減少反卷積(deconvolution)產生的環效應(ringing artifacts)，並帶來更好的去模糊效果。我們的演算法不需要任何外部的學習資訊，但可以達到和以機器學習為基礎的方法相似的效能並勝過其他單一影像盲反卷積的方法。Image deblurring in a fundamental problem in computer vision and image processing. The goal of image deblurring is to recover blur kernels and latent images from input blurred images. It is an inverse problem similar to image denoising and image super-resolution, but more challenging due to the diversity of motion blur kernels and natural images. Recent deblurring approaches include two main stages: blur kernel estimation and image restoration. Blur kernel estimation algorithms rely on sufficient significant edges on input images to infer a reliable blur kernel. Most approaches use only Y layer in the YUV color space to estimate blur kernels. However, researches show that RGB-to-Y conversion may lose contrast and edge information. In this thesis, we focus on solving single-image motion deblurring problem in which blurry images are produced by camera shakes with hand-held shots. A novel deblurring algorithm is proposed, which consider RGB color channels together in the blur kernel estimation stage. We also propose two color-image priors derived from the color-line model. Our color-image priors not only restore image contrasts around edges, but also reduce image noises caused by inaccurate estimated blur kernels during optimization. Experimental results show that our algorithm can enhance the stability and the accuracy of blur kernel estimation. Accurate blur kernels can reduce the ringing artifacts generated from deconvolution, and lead to better deblurred results. Our algorithm do not need any external learning information, but the performance of our algorithm is comparable to the learning-based method and outperforms other single-image blind deconvolution methods.口試委員會審定書 # 誌謝 i 中文摘要 iii ABSTRACT v CONTENTS vii LIST OF FIGURES xi LIST OF TABLES xvii Chapter 1 Introduction 1 1.1 Motivation 1 1.2 Image Deblurring Overview 1 1.3 Main Contribution 4 1.4 Organization 4 Chapter 2 Non-Blind Deconvolution 5 2.1 Direct Inverse Filter 6 2.2 Wiener Deconvolution 8 2.3 Richardson-Lucy Deconvolution 11 2.4 Regularized RL Algorithm 13 2.5 MAP-based algorithm 18 2.5.1 MAP Formulation 18 2.5.2 Fast Deconvolution using Hyper-Laplacian priors 25 2.5.3 Patch-based Method 30 2.6 Summary 31 Chapter 3 Blind Deconvolution 33 3.1 Hardware or Multiple Images Approaches 34 3.2 Single-Image Blind Deconvolution 36 3.2.1 Naive MAP Approaches 37 3.2.2 Extended MAP Approaches 40 3.2.3 Maximum Marginal Distribution Approaches 53 3.3 Summary 64 Chapter 4 Proposed Image Deblurring Method 65 4.1 Observation 65 4.2 Framework 68 4.3 The Color-Line Model 71 4.4 The Proposed Deblurring Algorithm 77 4.4.1 Solving Intermediate Latent Image 78 4.4.2 Solving Blur Kernel 81 4.4.3 Final Non-Blind Deconvolution 82 4.5 Experimental Results 83 4.5.1 Quantitative Comparison on Sun and Hays’ dataset [44] 84 4.5.2 Deblurring on Real-World Photos 98 4.6 Discussion 108 4.7 Summary 113 Chapter 5 Conclusion and Future Work 115 5.1 Conclusion 115 5.2 Future Work 116 Appendix A 117 Appendix B 123 REFERENCE 12712413704 bytesapplication/pdf論文公開時間：2017/07/11論文使用權限：同意無償授權影像去模糊最佳化逆問題盲反卷積問題點擴散函數/運動模糊函數估計色彩線模型thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/264263/1/ntu-103-R01942068-1.pdf