基于深度学习的表情识别

发布时间：2018-05-17 11:46

本文选题：人脸表情识别 + 深度学习　；参考：《南京邮电大学》2017年硕士论文

【摘要】：人脸表情能够表达人类的情绪、意图等,人脸表情识别作为情感智能系统的关键技术,是实现人机交互的重要基础。传统的表情识别中依靠人工精心设计的特征提取算法,不仅复杂,还会一定程度上丢失原有的表情特征信息。近年来,深度学习作为以纯数据为驱动的特征学习算法,能够自主地学习到样本的更加本质的特征,因此,本文将深度学习引入人脸表情识别任务中,探讨及研究深度学习在表情识别中的应用。本文的主要研究工作及成果总结如下:(1)扩增了人脸表情库。人脸表情库是表情识别的基本条件,本文对现有的表情库进行人脸检测、归一化等预处理,并采用数据集扩增策略扩增了人脸表情库。(2)研究了一种基于深度置信网络(Deep Belief Network,DBN)的人脸表情识别方法。通过预训练和微调阶段调整优化DBN模型参数,最顶层BP网络输出表情分类结果,在CK+数据库上取得了91.16%的识别率。(3)研究了一种基于卷积神经网络(Convolutional Neural Network,CNN)的人脸表情识别方法。人脸表情的变化往往是细微的,CNN可以捕捉到图像的局部特征,组合低层特征构成更加抽象的高层特征,从而更适合于二维表情图像的分类。相比DBN,基于CNN的人脸表情识别方法取得的识别率提高了5.02%。(4)研究了一种基于NIN(Network in Network)的人脸表情识别方法。先对低层特征局部进行更好地抽象有利于提升高层特征的表征能力,NIN的卷积层具有较强的非线性特征提取能力,因而有利于复杂的人脸表情图像的非线性特征的抽象和表达。相比CNN,基于NIN的人脸表情识别方法取得的识别率进一步提高了2.79%。(5)实现了一个人脸表情识别演示系统。该系统主要分为两个功能,一是人脸表情自动识别,将人脸表情分为七类;二是卷积层可视化,可以直观地观察到每个卷积层卷积运算后输出的特征图。
[Abstract]:Facial expression can express human emotion, intention and so on. As a key technology of emotional intelligence system, facial expression recognition is an important foundation of human-computer interaction. Traditional facial expression recognition based on artificial carefully designed feature extraction algorithm is not only complex, but also lose the original facial feature information to some extent. In recent years, as a feature learning algorithm driven by pure data, depth learning is able to learn more essential features of samples independently. Therefore, depth learning is introduced into facial expression recognition task in this paper. To explore and study the application of depth learning in facial expression recognition. The main work and results of this paper are summarized as follows: 1) the facial expression database is expanded. Facial expression database is the basic condition of facial expression recognition. An expression recognition method based on Deep Belief Network (DBN) is proposed. By adjusting the parameters of the DBN model in pre-training and fine-tuning stage, the top-level BP network outputs the facial expression classification results. The recognition rate of 91.16% is obtained on CK database.) A new facial expression recognition method based on Convolutional Neural Network (CNN) based on convolutional neural network is studied. The change of facial expression is usually subtle CNN can capture the local features of the image and combine the low-level features to form a more abstract high-level feature which is more suitable for the classification of two-dimensional facial expression images. Compared with DBN, the recognition rate of facial expression recognition method based on CNN is 5.022.The paper studies a face expression recognition method based on NIN(Network in Network. Firstly, the local abstraction of lower level features is helpful to enhance the representation ability of high-level features. The convolution layer of NIN has a strong ability to extract nonlinear features, which is conducive to the abstraction and expression of nonlinear features in complex facial expression images. Compared with the CNN-based facial expression recognition method, the recognition rate obtained by the NIN method is further improved by 2.79. 5) A human facial expression recognition demonstration system is implemented. The system is mainly divided into two functions, one is automatic facial expression recognition, facial expression is divided into seven categories, the other is convolution layer visualization, can directly observe each convolution layer after convolution output features.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】