在機(jī)器學(xué)習(xí)領(lǐng)域中普遍面臨處理大量且高維的多媒體數(shù)據(jù)問題。并且,如何從具有多樣性和非線性的多媒體數(shù)據(jù)中提取有效的鑒別性特征,是特征提取算法中具有挑戰(zhàn)性的課題。本文對以上問題進(jìn)行了研究,其核心思想是利用高維數(shù)據(jù)在實(shí)際應(yīng)用中往往具有低維的特點(diǎn),將數(shù)據(jù)的幾何結(jié)構(gòu)表示為流形圖結(jié)構(gòu)并進(jìn)行分析。論文具體介紹了三種新的多媒體數(shù)據(jù)分析方法,并取得了顯著的進(jìn)展。其中包括引入了多流形嵌入的字典誘導(dǎo)最小二乘框架,引入了圖嵌入的廣義多字典最小二乘框架,以及通過保持PCA框架的全局和局部結(jié)構(gòu)進(jìn)行流形對齊。第一種方法擴(kuò)展了主成分分析(PCA)的概念,通過最小化最小二乘重構(gòu)誤差思想保持?jǐn)?shù)據(jù)全局結(jié)構(gòu),并引入分布字典對丟失和噪聲數(shù)據(jù)點(diǎn)的離群分布對數(shù)據(jù)結(jié)構(gòu)重構(gòu)。接著,通過多流形嵌入保持純凈的局部結(jié)構(gòu)。因此,這種方法可以在低維投影中獲得鑒別信息,同時(shí)保持全局和局部結(jié)構(gòu)的平衡。我們提出的方法在多媒體數(shù)據(jù)分析方面進(jìn)行了大量實(shí)驗(yàn)并與目前最先進(jìn)方法相比表明該方法具有更好的性能。進(jìn)一步的,在此基礎(chǔ)上對第一種方法進(jìn)行擴(kuò)展,我們提出使用包含多個字典的第二種方法。在多視圖數(shù)據(jù)的情況下,多字典進(jìn)一步增強(qiáng)了對噪聲和冗余數(shù)據(jù)點(diǎn)的識別。接著,根據(jù)兩...

【文章頁數(shù)】：155 頁

【學(xué)位級別】：博士

【文章目錄】：
ABSTRACT
摘要
CHAPTER 1 INTRODUCTION
    1.1 BACKGROUND OF MULTIMEDIA DATA ANALYSIS
    1.2 THE STUDY SIGNIFICANCE
    1.3 CHALLENGES IN MULTIMEDIA DATA ANALYSIS
    1.4 CONTRIBUTIONS OF THE DISSERTATIQN
    1.5 ORGANIZATION OF THE DISSERTATION
CHAPTER 2 RELATED WORK
    2.1 REPRESENTATION LEARNING
        2.1.1 Dimensionality Reduction and Graph Embedding Techniques
            2.1.1.1 Global Structures Preserving Techniques
            2.1.1.2 Local Structures Preserving Techniques
            2.1.1.3 Multi-view Learning Techniques
        2.1.2 Manifold Alignment Techniques
CHAPTER 3 DICTIONARY-INDUCED LEAST SQUARES FRAMEWORK WITHMULTI-MANIFOLD EMBEDDINGS
    3.1 INTRODUCTION
    3.2 PROPOSED DLSME
        3.2.1 Achieving Lower Dimensions in DLSME
        3.2.2 Obtaining Parameter γ in our Proposed DLSME
    3.3 EXPERIMENTAL RESULTS
        3.3.1 Dataset Description
        3.3.2 Experimental Settings
        3.3.3 Results Discussions and Comparisons
            3.3.3.1 Comparison between DLS and PCA
            3.3.3.2 Web Image Annotation
            3.3.3.3 Visual Recognition
                (1) Digit Recognition
                (2) Object Recognition
            3.3.3.4 Computation Complexity
            3.3.3.5 Parameters α and r of DLSME
    3.4 SUMMARY
CHAPTER 4 A GENERALIZED MULTI-DICTIONARY LEST SQUARESFRAMEWORK REGULARIZED WITH MULTI-GRAPH EMBEDDINGS
    4.1 INTRODUCTION
    4.2 THE PROPOSED MD-MGE METHODS
        4.2.1 Obtaining Low Dimensional Projections in the Proposed Methods
        4.2.2 Obtaining α and β in the Proposed Methods
    4.3 EXPERIMENTAL RESULTS AND ANALYSIS
        4.3.1 Experimental Setting
        4.3.2 Handwritten Numerals Recognition
        4.3.3 Object Recognition
        4.3.4 Face Recognition
            4.3.4.1 Experiments on the ORL Dataset
            4.3.4.2 Experiments on the Extended YaleB Dataset
        4.3.5 Speech Recognition
        4.3.6 Computational Complexity and Time
        4.3.7 Control Parameters c & r of MD-MGE
        4.3.8 Comparison of MD-MGE and DLSME
        4.3.9 Evaluation of Experimental Results
    4.4 SUMMARY
CHAPTER 5 MANIFOLD ALIGNMENT VIA GLOBAL AND LOCAL STRUCTURESPRESERVING PCA FRAMEWORK
    5.1 INTRODUCTION
    5.2 THE PROPOSED MAPGL METHOD
        5.2.1 Global and Local Structures Preserving PCA Framework
        5.2.2 Optimizing MAPGL
        5.2.3 Obtaining Parameters γ and β in MAPGL
    5.3 EXPERPIMENTS
        5.3.1 Datasets Description
        5.3.2 Experimental Settings
        5.3.3 Alignment Experiments
            5.3.3.1 Protein Manifolds Alignment
            5.3.3.2 Rotated Objects Alignment
            5.3.3.3 Head Pose Images Alignment
            5.3.3.4 Image and Text Alignment
        5.3.4 Experiments on Visual Recognition
            5.3.4.1 Objects Recognition
                (i) Handwritten Numerals Recognition
                    (a) Experiments on MFD Dataset
                    (b) Experiments on USPS Dataset
                (ii) Face Recognition
                    (a) Experiments on YALE Dataset
                    (b) Experiments on AR Dataset
                    (c) Experiments on UMIST Dataset
        5.3.5 Effect of Parameters in MAPGL
    5.4 SUMMARY
CHAPTER 6 GENERAL CONCLUSIONS AND FUTURE WORK
    6.1 GENERAL CONCLUSIONS
    6.2 CONTRIBUTIONS
    6.3 FUTURE WORK
REFERENCES
ACKNOWLEDGEMENTS
PUBLICATIONS



本文編號：3958651