簡易檢索 / 詳目顯示

研究生: 桑卡拉斯里尼瓦桑·塞薩蒂裡
Sankarasrinivasan Seshathiri
論文名稱: 隨機分析與混合深度學習框架於半色調浮水印、分類及重建之應用
Stochastic Analysis and Hybrid Deep Learning Frameworks for Enhanced Halftone Watermarking, Classification and Reconstruction
指導教授: 郭景明
Jing-Ming Guo
口試委員: 郭景明
丁建均
楊家輝
王乃堅
貝蘇章
杭學鳴
鍾國亮
學位類別: 博士
Doctor
系所名稱: 電資學院 - 電機工程系
Department of Electrical Engineering
論文出版年: 2023
畢業學年度: 111
語文別: 英文
論文頁數: 147
中文關鍵詞: 數字半色調卷積神經網絡視覺加密反半色調擴散模型
外文關鍵詞: digital halftoning, convolutional neural networks, visual encryption, inverse halftoning, diffusion models
相關次數: 點閱:208下載:2
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報
  • 數字半色調是一種圖像轉換技術,可將灰度圖像轉換為相應的二值圖像,可用於打印。數字半色調研究領域具有廣泛的前景,包括水印、視覺加密、圖像質量評估、分類、逆半色調和圖像壓縮。本文提出了三個關鍵研究問題,重點是視覺加密、分類和逆半色調。
    第一個主題強調簇點半色調視覺加密方案的開發。所提出的方法涉及自適應抖動策略,該策略對相鄰和中間抖動陣列配置進行操作以獲得共軛抖動陣列對。該框架將秘密圖像嵌入到多個半色調圖像中,當主圖像相互重疊時,隱藏的信息可以很容易地被揭示。與現有方案相比,所提出的方法提供了更好的視覺質量、對比度和計算強度。
    此外,本研究引入了一種對半色調圖案進行分類的新方法,該方法可以推廣到所有半色調類型。所提出的方法利用隨機分析來揭示半色調圖案的獨特特徵。通過利用二元分佈和隨機幾何之間的內在聯繫,該方法提取隨機參數來構造手工向量。採用基於極限學習機的多分類器模型,實現快速、精準的分類。為了支持這一努力,我們創建了一個綜合數字數據庫,其中包含 96 個參考圖像並涵蓋 21 種不同的半色調和多色調類別。對所提出的方案的廣泛分析表明,其能夠對多種類型的半色調圖案實現 100% 的分類率。此外,還引入了一種新穎的自監督方法,專門用於滿足掃描圖像半色調分類背景下的實時要求。該方法利用自監督學習算法和增強增強策略來進一步提高分類性能。
    這項研究的結論部分集中於解決逆半色調問題,其中涉及將半色調圖像恢復為其原始版本。這個問題本質上是不適定的,對準確重建各種半色調圖案以及增強紋理和結構細節提出了挑戰。本文提出了兩種解決方案:提出了一種新穎的深度逆半色調方法,該方法結合了條件生成對抗網絡(C-GAN)、視覺變換模型(VT)和多同質分類器(VT-MHC) 。我們的方法引入了最小-最大損失函數,以及最小絕對偏差損失和結構相似性損失,以提高重建精度。實驗結果表明,與傳統和其他基於深度學習的技術相比,具有卓越的泛化性和性能。我們的集成方法克服了以前方法的局限性,實現了高質量的重建。此外,我們提出了一種使用自監督深度學習方法進行重建的替代方法。這種方法利用 Cycle-GAN 架構和圖像到圖像擴散模型,實現與監督方法相當的一致性能。


    Digital halftoning is an image transformation technique to convert the gray scale image into its corresponding binary images, which can be useful for printing. The field of research in digital halftoning offers a wide range of prospects, including watermarking, visual encryption, image quality assessment, classification, inverse halftoning, and image compression. In this thesis, three key research problems are presented, focusing on visual encryption, classification, and inverse halftoning.
    The first topic emphasizes on the development of visual encryption scheme for the clustered dot halftones. The proposed approach involves adaptive dithering strategy which operates on the adjacent and intermediate dither array configuration to arrive at the conjugate dither array pairs. The framework embeds the secret image into multiple halftone images and when the host images are overlaid on each other, the hidden message can be easily revealed. In comparison with the existing schemes, the proposed approach offers better visual quality, contrast and computationally less intensive.
    Furthermore, this research introduces a novel approach to classifying halftone patterns which can be generalized to all halftone types. The proposed approach capitalizes on stochastic analysis to uncover distinct characteristics of halftone patterns. By leveraging the inherent connection between binary distribution and stochastic geometry, the method extracts stochastic parameters to construct hand-crafted vectors. A multi-classifier model, specifically based on extreme learning machine, is employed to achieve rapid and precise classification. To support this endeavor, a comprehensive digital database comprising 96 reference images and encompassing 21 different varieties of halftone and multitone classes has been created. Extensive analysis of the proposed scheme demonstrates its capability to achieve a 100% classification rate for numerous types of halftone patterns. Further, a novel self-supervised approach is also introduced, specifically designed to address real-time requirements in the context of halftone classification for scanned images. The proposed method utilizes self-supervised learning algorithms and enhanced augmentation strategies to further improve the classification performance.
    The concluding aspect of this research concentrates on addressing the inverse halftoning problem, which involves the restoration of halftone images to its original version. This problem is inherently ill-posed, posing challenges in accurately reconstructing various halftone patterns and enhancing texture and structural details. Two solutions are proposed in this thesis: A novel deep inverse halftoning method that combines conditional generative adversarial nets (C-GANs), a vision transformer model (VT), and a multi-homogeneity classifier (VT-MHC) is proposed. Our approach introduces the min-max loss function, along with the least absolute deviation loss and structural similarity loss, to improve reconstruction accuracy. Experimental results demonstrate superior generalization and performance compared to conventional and other deep learning-based techniques. Our integrated approach overcomes the limitations of previous methods and achieves high-quality reconstruction. Furthermore, we present an alternative approach for reconstruction using a self-supervised deep-learning method. This approach leverages the Cycle-GAN architecture and image-to-image diffusion models, enabling consistent performance comparable to supervised approaches.

    Recommendation Letter ……………………………………………………………………II Approval Letter ……………………………………………………………………………III Abstract IV Acknowledgment VI Table of Contents VII List of Figures X List of Tables XIII List of Abbreviations XV Chapter 1 Introduction to Digital Halftoning 16 1.1 Introduction 16 1.2 Halftone Research Prospects 20 1.2 Motivation of the Thesis 27 1.3 Key Objectives 28 1.4 Thesis Organization 30 Chapter 2 Visual Encryption Scheme for Clustered-Dot Halftones 31 2.1 Literature Review 32 2.1 Key Limitation and Objectives 35 2.2 Proposed Visual Encryption Scheme 37 2.2.1 Screen Generation for Clustered-Dot Halftones 37 2.2.2 Proposed Visual Encryption Scheme. 41 2.3 Results and Discussion 45 Summary 56 Chapter 3 Unified and Self-Supervised Halftone Classification using Stochastic Analysis 57 3.1 Halftone Types 57 3.1.1 Ordered Dithering 57 3.1.2 Error Diffusion 58 3.1.3 Dot Diffusion 59 3.1.4 Iterative Approaches 60 3.2 Literature Review 61 3.3 Limitation and Contribution 61 3.4 Digital Halftone Database (DHD) 62 3.5 Proposed Halftone Classification Method 65 3.5.1 Stochastic Halftone Analysis 65 3.5.2 Steps for stochastic halftone model training 69 3.6 Experimental Results 72 3.7 Scanned Halftone Classification using Self-Supervised Learning Models 76 3.7.1 Proposed SSL Halftone Classification Approach 77 Summary 91 Chapter 4 Supervised & Unsupervised Deep Models for Generalized Inverse Halftoning 92 4.1 Literature Review 93 4.2 Proposed Self -Supervised Learning Model 97 4.2.1 Limitation and Objectives 97 4.2.2 Proposed Halftone Reconstruction Model. 99 4.2.3 Homogeneity Analysis 109 4.2.4 Results and Discussion 112 4.2 Self-Supervised Halftone Reconstruction 122 4.2.1 Cycle-GAN for Halftone 122 4.2.2 Image Refinement using diffusion model 126 4.2.3 Results and Discussion 127 Summary 129 Conclusion 131 Future Scope 132 Reference 134 List of Publication 144

    數字半色調是一種圖像轉換技術,可將灰度圖像轉換為相應的二值圖像,可用於打印。數字半色調研究領域具有廣泛的前景,包括水印、視覺加密、圖像質量評估、分類、逆半色調和圖像壓縮。本文提出了三個關鍵研究問題,重點是視覺加密、分類和逆半色調。
    第一個主題強調簇點半色調視覺加密方案的開發。所提出的方法涉及自適應抖動策略,該策略對相鄰和中間抖動陣列配置進行操作以獲得共軛抖動陣列對。該框架將秘密圖像嵌入到多個半色調圖像中,當主圖像相互重疊時,隱藏的信息可以很容易地被揭示。與現有方案相比,所提出的方法提供了更好的視覺質量、對比度和計算強度。
    此外,本研究引入了一種對半色調圖案進行分類的新方法,該方法可以推廣到所有半色調類型。所提出的方法利用隨機分析來揭示半色調圖案的獨特特徵。通過利用二元分佈和隨機幾何之間的內在聯繫,該方法提取隨機參數來構造手工向量。採用基於極限學習機的多分類器模型,實現快速、精準的分類。為了支持這一努力,我們創建了一個綜合數字數據庫,其中包含 96 個參考圖像並涵蓋 21 種不同的半色調和多色調類別。對所提出的方案的廣泛分析表明,其能夠對多種類型的半色調圖案實現 100% 的分類率。此外,還引入了一種新穎的自監督方法,專門用於滿足掃描圖像半色調分類背景下的實時要求。該方法利用自監督學習算法和增強增強策略來進一步提高分類性能。
    這項研究的結論部分集中於解決逆半色調問題,其中涉及將半色調圖像恢復為其原始版本。這個問題本質上是不適定的,對準確重建各種半色調圖案以及增強紋理和結構細節提出了挑戰。本文提出了兩種解決方案:提出了一種新穎的深度逆半色調方法,該方法結合了條件生成對抗網絡(C-GAN)、視覺變換模型(VT)和多同質分類器(VT-MHC) 。我們的方法引入了最小-最大損失函數,以及最小絕對偏差損失和結構相似性損失,以提高重建精度。實驗結果表明,與傳統和其他基於深度學習的技術相比,具有卓越的泛化性和性能。我們的集成方法克服了以前方法的局限性,實現了高質量的重建。此外,我們提出了一種使用自監督深度學習方法進行重建的替代方法。這種方法利用 Cycle-GAN 架構和圖像到圖像擴散模型,實現與監督方法相當的一致性能。

    無法下載圖示
    全文公開日期 2028/07/31 (校外網路)
    全文公開日期 2028/07/31 (國家圖書館:臺灣博碩士論文系統)
    QR CODE