指導教授：丁建均臺灣大學：電信工程學研究所陳信慧Chen, Hsin-HuiHsin-HuiChen2014-11-302018-07-052014-11-302018-07-052014http://ntur.lib.ntu.edu.tw//handle/246246/264275隨著多媒體技術的進步，大量的資料如影像和視訊等資料需要加以壓縮以便於儲存和傳輸。因此，資料壓縮是一項重要的研究議題。雖然目前有許多不同的方法已經提出解決方案，我們相信其仍有改善的空間。另一方面，影像還原在電腦視覺和影像處理領域已經成為一個很熱門的研究方向。在傳統的影像濾波方法中，常利用在影像內容中分布於鄰近區域的資料來計算最佳化的權重值。不過，這些方法都忽略了高重複性的影像結構特性，因而限制了整體效果。自從非區域性平均去雜訊演算法被提出來，在近年來許多很有效的影像去雜訊、影像內插和影像去模糊方法以此為基礎陸續的發展出來。基於此非區域性原則，我們可以充分利用在影像中的非區域自相似性區塊特徵來開發影像還原演算法。 可適性預測方法在失真/無失真影像壓縮和影像還原扮演一個關鍵的角色。在影像壓縮中，我們研究目標包含移除影像中的空間冗餘性、將Z字形係數掃描法一般化，或是計算最佳參數來得出較短的編碼長度。在影像還原研究中，目標為內插出新的像素值以增加影像解析度或是去除影像中的雜訊。為了達成以上的研究目標，需要針對不同的情況來適當地設計可適性預測方法。 在本論文中，我們採取區域性，非區域性和全域性預測來改善無失真和失真影像壓縮演算法，以及在影像還原議題中影像去雜訊和影像內插的方法。本論文總共包含五個獨立的研究成果。在失真影像研究中，我們提出了兩個研究成果。其中一項為以聯合機率為基礎的可適性格倫布編碼演算法。參考欲編碼像素之鄰近區域的資訊，動態地調整格倫布參數來編碼影像，以產生較短的編碼長度。另一項成果為針對使用離散餘弦轉換為主的影像壓縮系統，提出以區域性和全域性預測為基礎的可適性係數掃描法，將Z字形掃描法一般化。參考來自之前編碼過的影像區塊，利用區域性和全域性資訊來提升整體壓縮效能。在無失真影像研究中，為了降低預測影像誤差值以及估測其精確的機率分布，我們提出了非區域性內容模型和可適性預測方法來降低位元率。此方法主要的精神為利用在影像中非區域結構的自相似性，而此特性常出現在影像空間和預測誤差空間。 針對影像去雜訊的研究，我們提出了一個強很有效的方法，稱為以雙向主成分分析為基礎的非區域性平均之影像去雜訊演算法。更進一步將其實現在一種不需迭代計算的粗略至精細演算法，無論影像受到低雜訊或高雜訊干擾的情況下，去雜訊後的影像皆能保留邊緣或紋理的完整性。而在影像內插的研究中，我們將影像結構的自相似性度量方法應用在非區域性邊緣導向的影像內插方法中，如此得到較佳的影像品質。整體而言，本論文中此五項主要的研究方法皆充分地利用區域性，非區域性或全域性資訊來提升整體的影像壓縮和影像還原效果。With the advancement of multimedia technologies, the huge amount of data such as images and videos, etc., needs to be compressed for storage and transmission. There-fore, data compression is a very important issue. Although many different methods have been proposed, it is believed that there is still room for improvement. On the other hand, image restoration has been an active research topic in computer vision and image processing. In conventional filtering methods, the local neighboring image data is usually used for weight optimization. It limits the overall performance due to ignoring the repetitive structure patterns in images. Since the work of the non-local means (NLM) algorithm, many more powerful methods for image denoising, interpola-tion and deblurring are proposed in recent years based on the non-local principle, which exploits the non-local self-similarity between the patches in an image. Adaptive prediction is a key component in lossy/lossless image compression and image restoration. To remove the spatial redundancy in images, the adaptive prediction methods for generalizing the zigzag scanning method, obtaining the better coding pa-rameter for shorter codelength in image compression, interpolating missing pixels or filtering out noise in image restoration are developed. In this dissertation, I adopt the idea of local, non-local, and global predictions to improve lossless and lossy image compression methods and image restoration methods, such as image denoising and image interpolation. In sum, there are five individual re-search works. For lossy image coding, two research works are presented. One is the proposed joint-probability-based adaptive Golomb coding (JPBAGC) algorithm, which takes the local neighboring data into account to adaptively adjust the Golomb parameter for yielding shorter codelength. The other one is the local- and global-prediction-based adaptive scanning (LGPAS), which generalizes the zigzag scanning method. It is pro-posed to achieve a better compression performance using the local and global infor-mation from previously encoded/decoded blocks in DCT-based image coding systems. For lossless image coding, non-local context modeling and adaptive prediction (NCMAP) are proposed to reduce the image prediction errors and estimate their true probability distribution by exploiting the non-local structural self-similarity in the spatial and prediction error domain. For image denoising, a powerful and efficient scheme, called non-local means based on bidirectional principal component analysis (NLM-BDPCA), is proposed. Fur-thermore, the coarse-to-fine algorithm is also implemented. It does not perform de-noising iteratively and can well preserve the edge/texture information. For image inter-polation, the structural similarity based metric is incorporated into the framework of non-local edge-directed image interpolation (SSNLEDI) for yielding better peak-to-signal ratio (PSNR) values. Overall, these five efforts take the advantage of the rich information extracted from local, non-local, or global regions in images to achieve better performance for image compression and image restoration.誌謝 ii 中文摘要 iii ABSTRACT v CONTENTS vii LIST OF FIGURES vii LIST OF TABLES xix Chapter 1 Introduction 1 1.1 Background and Motivation 1 1.2 Main Contributions 5 1.3 Organization 8 Chapter 2 Overview of Image Compression 9 2.1 JPEG 9 2.2 The H.264-Intra JPEG-based Image Coding System 10 2.3 Review of Golomb Codes 11 2.4 Context-based Predictive Lossless Image Coding 13 2.4.1 Gradient-Adjusted Prediction (GAP) 14 2.4.2 Edge-Directed Prediction (EDP) 16 2.5 Entropy-Coding Analysis of Context Modeling 17 Chapter 3 Overview of Image Restoration 20 3.1 Overview of Image Denoising 20 3.1.1 Classical Gaussian Filter 21 3.1.2 Bilateral Filter 22 3.1.3 Non-local Means Based on Structure and Homogeneous Patch Simi-larity (NLM-SHPS) 22 3.1.4 Non-local Means (NLM) 23 3.1.5 Principal Neighborhood Dictionaries for Non-local Means Image De-noising 24 3.2 Overview of Image Interpolation 26 3.2.1 New Edge-Directed Interpolation (NEDI) 26 3.2.2 Non-local Edge-Directed Interpolation (NLEDI) 29 3.2.3 Soft-Decision Interpolation (SAI) 29 3.2.4 Robust Soft-Decision Interpolation Using Weighted Least Squares (RSAI) 32 3.2.5 Bilateral Soft-Decision Interpolation for Real-Time Applications (BSAI) 32 Chapter 4 Adaptive Golomb Code for Joint Geometrically Distributed Data and Its Application in Image Coding 35 4.1 Introduction 36 4.2 Proposed Joint-Probability-based Adaptive Golomb Codes 38 4.3 Using The Proposed Joint-Probability-Based Adaptive Golomb Codes to Improve JPEG-Based Image Coding 46 4.3.1 Encoding The Difference in DC Terms Using The Proposed JPBAGC 47 4.3.2 Encoding The AC Terms Using The Proposed JPBAGC 48 4.3.3 Overall Performance 51 4.4 Using The Proposed JPBAGC To Improve The H.264-Intra JPEG-Based Image Coding System 55 4.5 Conclusions 58 Chapter 5 Non-local Context Modeling and Adaptive Prediction for Lossless Image Coding 60 5.1 Introduction 60 5.2 Proposed Adaptive Prediction 63 5.2.1 Weighted Edge-Directed Prediction (WEDP) 63 5.2.2 Non-local Predictor (NLP) 64 5.2.3 Proposed Adaptive Prediction 65 5.3 Proposed Non-local Context Modeling and Context Generation 65 5.3.1 Non-local Context Modeling 66 5.3.2 Context Generation and Entropy Coding 69 5.4 Simulation Results 69 5.5 Conclusions 70 Chapter 6 New Adaptive Coefficient Scanning based on Local and Global Predictions for JPEG and H.264/AVC Intra-Coding 72 6.1 Introduction 73 6.2 Proposed LGPAS Scheme for JPEG 75 6.2.1 Local Prediction to Estimate the Variance of the DCT Coefficients 75 6.2.2 Local Prediction to Estimate the Probability of nonzeo DCT Coefficients 76 6.2.3 Two Weighting Matrices: Zigzag Weighting Matrix and Weighting Matrix 76 6.2.4 Scan-Order Pattern Generation 77 6.3 Proposed LGPAS Scheme for H.264-Intra JPEG 78 6.3.1 Local Prediction to Estimate the Variances of Probabilities of the Nonzero DCT Coefficients 78 6.3.2 Global Prediction to Estimate the Variances and Probabilites of the Nonzero DCT Coefficients 79 6.3.3 Scan-Order Pattern Generation 80 6.4 Simulation Results 80 6.5 Conclusions 83 Chapter 7 Non-local Means Image Denoising Based on Bidirectional Principal Component Analysis 89 7.1 Introduction 90 7.2 Proposed Non-local Means Based on Bidirectional Principal Component Analysis (NLM-BDPCA) 92 7.2.1 Bidirectional Principal Component Analysis 92 7.2.2 Coarse-to-Fine NLM-BDPCA Image Denoising 96 7.3 Simulation Results 98 7.4 Conclusions 107 Chapter 8 Structural Similarity-based Non-local Edge-Directed Image Interpolation 108 8.1 Introduction 108 8.2 Proposed Strucutural Similarity-based NLEDI 111 8.2.1 The SSIM-based Metric for Weighted Averaging 112 8.2.2 Generalizations to Arbitrary Upscaling Factor 115 8.3 Simulation Results 116 8.4 Conclusions 117 Chapter 9 Conclusions 119 REFERENCE 1225420578 bytesapplication/pdf論文公開時間：2017/07/29論文使用權限：同意有償授權(權利金給回饋學校)影像壓縮霍夫曼編碼法格倫布編碼法可適性格倫布編碼法鋸齒狀掃描可適性係數掃描法H.264/AVC畫面內預測法邊緣導向預測法非區域性平均法結構自相似性雙向主成分分析法影像去雜訊影像內插thesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/264275/1/ntu-103-D98942018-1.pdf