基于深度压缩时空模型的视频表情识别在边缘设备的实现

发布时间：2021-12-29 08:20

　　近年来深度学习逐渐成为信息科学领域的研究热点,而随着基于深度学习方法的研究技术的不断推进,数据特征信息的提取和处理效率获得了极大的提升,同时也推动了深度学习在计算机视觉、语音处理和自然语言处理等相关领域的迅猛发展。作为计算机视觉领域中一个比较重要的研究子方向,人脸表情识别可以广泛地应用到多个领域如人机交互、不良状态检测等。通常来说,表情不仅是一种非语言交际的方式,可以传递用于交流的辅助信息,也是人类情绪精神状态的潜在反映。通过表情辅助消息传递,可以让消息的信息量更为丰富,而消息接收者也能更为准确地把握信息的特征。所以在人机交互方面,表情识别可以用于让机器更准确地获取用户传递的消息内容;而在一些需要判断用户状态的场景,也可以利用表情识别完成不良状态的识别,比如判断驾驶员是否处于疲劳驾驶状态。通常非深度学习方法的表情识别主要是利用人工选取的表情特征,包括几何特征、统计特征和运动特征等,以及分类判别器的决策分类进行表情识别。这些方法都取得了一定的效果,但是过度依赖于特征的人工选取,鲁棒性较差,同时计算量非常大。深度学习方法则避免了特征的人工选取,同时其数据冗余度也保证了表情识别系统的鲁棒性。... 

【文章来源】：哈尔滨工业大学黑龙江省 211工程院校 985工程院校

【文章页数】：113 页

【学位级别】：硕士

【文章目录】：
详细中文摘要
ABSTRACT
CHAPTER 1. INTRODUCTION
    1.1 RESEARCH BACKGROUND AND SIGNIFICANCE
    1.2 RESEARCH STATUS OF RELATED FIELDS
    1.3 MAIN RESEARCH CONTENTS OF THIS SUBJECT
CHAPTER 2. FOUNDATIONS OF CONVOLUTIONAL NEURAL NETWORK ANDOPTIMIZATION
    2.1 INTRODUCTION
    2.2 COMPUTATIONS IN CONVOLUTIONAL NEURAL NETWORK
    2.3 GENERAL MATRIX MULTIPLICATION ALGORITHM FOR CNNACCELERATION
    2.4 WINOGRAD ALGORITHM FOR CONVOLUTIONAL LAYERACCELERATION
    2.5 BRIEF SUMMARY
CHAPTER 3. COMPRESSION AND OPTIMIZATION FOR LONG SHORT-TERMMEMORY
    3.1 INTRODUCTION
    3.2 COMPUTATIONS IN LONG SHORT-TERM MEMORY
    3.3 CLASSICAL DECOMPOSITION METHODS
        3.3.1 Tensor basics
        3.3.2 Singular value decomposition
        3.3.3 Tucker decomposition
    3.4 TENSORIZED COMPRESSION FOR LSTM ACCELERATION
        3.4.1 Tensor train tensor decomposition
        3.4.2 Tensorized compression for LSTM
    3.5 BRIEF SUMMARY
CHAPTER 4. SPATIOTEMPORAL MODEL FOR VIDEO FACIAL EXPRESSIONRECOGNITION
    4.1 INTRODUCTION
    4.2 FEATURE EXTRACTION WITH CNN
    4.3 SPATIOTEMPORAL LSTM MODEL
    4.4 OUR PROPOSED FRAMEWORK FOR FER
    4.5 BRIEF SUMMARY
CHAPTER 5. IMPLEMENTATION ON EDGE DEVICES
    5.1 INTRODUCTION
    5.2 IMPLEMENTATION ON ARM CPU
        5.2.1 ARM architecture and NEON technology
        5.2.2 ARM intrinsic programming
        5.2.3 Workflow overview on ARM CPU
    5.3 IMPLEMENTATION ON EDGE DEVICE 1
        5.3.1 Introduction to RK3399 Pro board
        5.3.2 Neural process unit and systolic arrays
        5.3.3 Workflow overview on RK3399 Pro board
    5.4 IMPLEMENTATION ON EDGE DEVICE 2
        5.4.1 Introduction to Atlas 200 Developer Kit
        5.4.2 Ascend 310 chipset and Da Vinci core
        5.4.3 Workflow overview on Atlas 200 DK board
CHAPTER 6. EXPERIMENTAL RESULTS AND PERFORMANCEANALYSIS
    6.1 INTRODUCTION
    6.2 EVALUATION ON FACIAL EXPRESSION RECOGNITIONDATASETS
    6.3 PERFORMANCE ANALYSIS
CHAPTER 7. CONCLUSIONS
结论
REFERENCES
PAPERS PUBLISHED IN THE PERIOD OF MASTER EDUCATION
ACKNOWLEDGEMENT



本文编号：3555752