近年來,深度學(xué)習(xí)方法極大地提高了人臉識別的準(zhǔn)確性,為了獲得更高的識別準(zhǔn)確率,集成學(xué)習(xí)可以應(yīng)用于深度學(xué)習(xí)算法中。傳統(tǒng)識別算法難以捕捉到面部表情所傳遞的有用信息,面部表情識別存在分辨率低、遮擋、光照、位置等問題,通常情況下,由于這些面部表情分類很差,人類無法識別它們。此外,面部表情的分類比較特殊,例如面部微笑并不總是意味著開心,面部表情往往取決于文化。然而,提高面部表情識別準(zhǔn)確率可以應(yīng)用到更靈敏、更智能的系統(tǒng),從而改善用戶體驗(yàn)。為了提高分類器的性能,降低人臉表情識別的錯(cuò)誤率,研究者開展了很多的工作,例如基于深度學(xué)習(xí)方法。有時(shí)候深度學(xué)習(xí)對面部表情識別存在困難,原因有很多,比如基于深度學(xué)習(xí)人臉面部表情識別應(yīng)用是一項(xiàng)復(fù)雜而困難的任務(wù),又例如很難找到高質(zhì)量的數(shù)據(jù)集,深度網(wǎng)絡(luò)的性能在很大程度上依賴于大量的標(biāo)記樣本。本文提出了一種基于卷積神經(jīng)網(wǎng)絡(luò)和集成深度網(wǎng)絡(luò)的新方法,可面向小樣本數(shù)據(jù)集分類情況,這些方法分別是多視角卷積神經(jīng)網(wǎng)絡(luò)(MVCNN)和集成遷移學(xué)習(xí)網(wǎng)絡(luò)(ETLN)。首先,將人臉圖像通過不同尺度進(jìn)行下采樣,然后向上采樣到統(tǒng)一圖像大小,得到多視角訓(xùn)練樣本。然后,構(gòu)造了一個(gè)具有雙通道特征提取結(jié)構(gòu)的多...

【文章頁數(shù)】：85 頁

【學(xué)位級別】：碩士

【文章目錄】：
ABSTRACT
摘要
List of Symbols
List of Abbreviations
Chapter 1 Introduction
    1.1 Background
    1.2 Motivation of Our Work
    1.3 Structure of The Thesis
Chapter 2 Related Works
    2.1 Facial Expression Recognition
    2.2 Literature Review
    2.3 Neuron Model
    2.4 Summary
Chapter 3 Multi-view Network based on CNN
    3.1 Convolutional Neural Networks (CNN)
    3.2 Multi-view CNN
        3.2.1 Multiple View Datasets
        3.2.2 Convolutional Layer
        3.2.3 Pooling Layer
        3.2.4 Fully Connected Layer
        3.2.5 Batch Normalization Layer
        3.2.6 Softmax Layer
        3.2.7 Pre-Processing
        3.2.8 Network Training
    3.3 Datasets
        3.3.1 The FER2013 Dataset
        3.3.2 The RAF-BASIC Dataset
    3.4 Results on FER2013 and Discussions
        3.4.1 Experimental Condition
        3.4.2 Results of DCNN with no data Aug
        3.4.3 Results of DCNN with data Aug
        3.4.4 Results of Multi-view CNN
    3.5 Results on RAF-BASIC and Discussions
        3.5.1 Results of DCNN with data Aug
        3.5.2 Results of Transfer DCNN
    3.6 Performance Evaluation of MVCNN and Transfer DCNN
    3.7 Summary
Chapter 4 Ensemble Transfer Learning Network (ETLN)
    4.1 Feature Learning
        4.1.1 VGG16
        4.1.2 VGG-face
        4.1.3 Ensemble and Transfer Learning
        4.1.4 Pre-Processing and Training Process
    4.2 Experimental Details
        4.2.1 Experimental Results on FER2013
        4.2.2 Experimental Results on RAF-BASIC
    4.3 Weights Analysis
    4.4 Special Combination
    4.5 Evaluation of The Proposed ETLN
    4.6 Summary
Chapter 5 Conclusions and Future Work
    5.1 Conclusions
    5.2 Future Work
References
Acknowledgements
Biography



本文編號：3909618