深度學(xué)習(xí)（DL）作為規(guī)模圖像大數(shù)據(jù)的聚類與分類的一種有效工具,展現(xiàn)了其解決無監(jiān)督和半監(jiān)督條件下的聚類與分類算法表示問題的無限潛能。卷積神經(jīng)網(wǎng)絡(luò)如今被運(yùn)用在許多先進(jìn)的聚類分類算法中。然而,隨著數(shù)字圖像的不斷增長,冗余的、無關(guān)的噪聲樣本也隨之增多,從而導(dǎo)致了卷積神經(jīng)網(wǎng)絡(luò)的性能的逐漸下降。卷積神經(jīng)網(wǎng)絡(luò)需要大量的有類標(biāo)的樣本來訓(xùn)練網(wǎng)絡(luò)。然而有類標(biāo)的數(shù)據(jù)常常較難獲得,價(jià)格昂貴且耗時。為了克服這個問題,我們該算法將卷積神經(jīng)網(wǎng)絡(luò)和模糊粗糙理論相結(jié)合提出了一種有效的針對規(guī)模圖像大數(shù)據(jù)的無監(jiān)督和半監(jiān)督算法,這其中主要的概念是降低模糊和粗糙的不確定性,同時更特別的是利用神經(jīng)網(wǎng)絡(luò)從原始數(shù)據(jù)中去掉噪聲樣本。本文提出的算法如下:1.首先,我們提出一種無監(jiān)督的聚類算法（FRCNN）。中心思想是利用多層卷積神經(jīng)網(wǎng)絡(luò)來學(xué)習(xí)高緯表達(dá)式,通過神經(jīng)網(wǎng)絡(luò)中的一個聚類層來初始化聚類中心。在訓(xùn)練過程中圖像的聚類中心和表達(dá)式交替更新。FRCM用來在前向傳播過程中更新聚類中心,同時利用基于隨機(jī)梯度下降法的后向傳播來更新神經(jīng)網(wǎng)絡(luò)權(quán)值。2.其次,本文構(gòu)造了一種半監(jiān)督模糊粗糙卷積神經(jīng)網(wǎng)絡(luò)（SSFRCNN）,將模糊粗糙C均值聚類和卷積神經(jīng)網(wǎng)... 

【文章來源】：西安電子科技大學(xué)陜西省 211工程院校 教育部直屬院校

【文章頁數(shù)】：145 頁

【學(xué)位級別】：博士

【文章目錄】：
摘要
ABSTRACT
List of Abbreviations
Chapter 1 Introduction
    1.1. Rough Set Theory (RST)
    1.2. Fuzzy C-Mean Clustering (FCM)
    1.3. Fuzzy Rough C-Mean Clustering (FRCM)
    1.4. Rough Set Attributes Reduction (RSAR)
    1.5. Convolution Neural Network
    1.6. Motivations
        1.6.1. Motivation of Fuzzy Rough C-Mean Based Unsupervised CNN Clustering forLarge-Scale Image Data
        1.6.2. Motivation of the A Semi-Supervised CNN with Fuzzy Rough C-Mean forImage Classification
        1.6.3. Motivation of Rough-KNN Noise-Filtered Convolutional Neural Network forImage Classification
        1.6.4. Motivation of Rough Noise-Filtered Easy Ensemble for Software FaultPrediction
    1.7. Contributions and Organization
        1.7.1. Contribution of Fuzzy Rough C-Mean Based Unsupervised CNN Clusteringfor Large-Scale Image Data
        1.7.2. Contribution of a Semi-Supervised CNN with Fuzzy Rough C-Mean for ImageClassification
        1.7.3. Contribution of Rough-KNN Noise-Filtered Convolutional Neural Networkfor Image Classification
        1.7.4. Contribution of Rough Noise-Filtered Easy Ensemble for Software FaultPrediction
        1.7.5. Organization
Chapter 2 Fuzzy Rough C-Mean Based Unsupervised CNN Clustering for Large-ScaleImage Data
    2.1. Introduction
    2.2. Fuzzy Rough C-Mean Based Unsupervised CNN Clustering
        2.2.1. The Problem of Deep-Learning-Based Clustering
        2.2.2. Background of Fuzzy Rough C-Mean (FRCM)
        2.2.3. Theoretical description of Proposed Approach
            2.2.3.1. FRUCNN Clustering Architecture
            2.2.3.2. Joint Clustering and Representation Learning
                2.2.3.2.1. Pre-Processing Data for UCNN
                2.2.3.2.2. Cluster Centroid Updating
                2.2.3.2.3. Representation Learning
    2.3. Experiments
        2.3.1. Data Preparation
        2.3.2. Performance Measure
        2.3.3. Comparison Schemes
        2.3.4. Implementation Details
        2.3.5. Experimental Design
            2.3.5.1. Computational Time Comparison
            2.3.5.2. Performance on Number of Cluster (k)
            2.3.5.3. Performance on Number of Epochs
    2.4. Threats to Validity
    2.5. Chapter Summary
Chapter 3 A Semi-Supervised CNN with Fuzzy Rough C-Mean for Image Classification
    3.1. Introduction
    3.2. A Semi-Supervised Fuzzy Rough Convolutional Neural Network (SSFRCNN)
        3.2.1. Framework of Our Approach
        3.2.2. Theoretical description of Proposed Approach
        3.2.3. Semi-Supervised Fuzzy Rough Convolution Neural Network (FRCNN)Training
        3.2.4. Mathematical Description
    3.3. Experiments
        3.3.1. Data Preparation
        3.3.2. Experimental Setup
        3.3.3. Experiment Result and Analysis
        3.3.4. Time Complexity
        3.3.5. Convergence Analysis of Semi-Supervised Fuzzy Rough ConvolutionalNeural Network (SSFRCNN)
    3.4. Chapter Summary
Chapter 4 Rough-KNN Noise-Filtered Convolutional Neural Network for ImageClassification
    4.1. Introduction
    4.2. Rough Set Theory Based 2d-Reduction Method
        4.2.1. Framework of Our Approach
        4.2.2. Theoretical description of Proposed Approach
    4.3. Experiments
        4.3.1. Data Preparation
        4.3.2. Implementation of Experiment
        4.3.3. Experiment Design&Analysis
            4.3.3.1. MNIST
            4.3.3.2. CIFAR-10
            4.3.3.3. YTF (Youtube-Face)
    4.4. Chapter Summary
Chapter 5 Rough Noise-Filtered Easy Ensemble for Software Fault Prediction
    5.1. Introduction
    5.2. Rough Noise-Filtered Easy Ensemble for Software Fault Prediction
        5.2.1. Framework of Our Approach
        5.2.2. Theoretical description of Proposed Approach
            5.2.2.1. Information Gain (IG) based Feature Selection
                5.2.2.1.1. Information Gain (IG)
                5.2.2.1.2. Symmetrical Uncertainty (SU)
            5.2.2.2. Rough- KNN Noise Filter (RK- Filter)
            5.2.2.3. Rough-KNN Noise Filtered Easy Ensemble (RKEE)
    5.3. Experiments
        5.3.1. Data Preparation
        5.3.2. Performance Measure
        5.3.3. Classification Models
        5.3.4. Experimental Design
        5.3.5. Result and Analysis
            5.3.5.1. Analysis of X-All verses X
            5.3.5.2. The effectiveness of Noise-Filter through KNN rule
            5.3.5.3. Impact of Rough set theory with Noise-Filter
            5.3.5.4. The impact of the combination of feature selection with Rough Noise-FilterEasy Ensemble
            5.3.5.5. Relationship between the performance and imbalanced ratio
            5.3.5.6. Comparison of different Schemes
    5.4. Chapter Summery
Chapter 6 Concluding Remarks and Future Work
    6.1. Concluding Remarks
    6.2. Future work
References
Acknowledgements
Bibliography



本文編號：3368289