目录
摘要 III
关键字 III
ABSTRACT IV
KEY WORDS IV
1 绪论 １
1.1 研究背景和意义 １
1.1.1 人脸表情识别的重要性 １
1.1.2 标记分布学习在人脸表情识别中的适用性 １
1.2 国内外研究现状 ２
1.2.1 人脸表情识别研究现状 ２
1.2.2 标记分布学习研究现状 ３
1.3 本文主要研究内容 ３
2 标记分布学习理论 ４
2.1 标记分布学习 ４
2.2 表情识别性能评价指标 ５
3 标记分布学习算法 ６
3.1 LDLSCL算法 ６
3.2 LDLSF算法 ８
3.3 EDLLRL算法 １０
4 参数优化 １１
4.1 LDLSCL优化 １１
4.2 LDLSF优化 １４
4.3 EDLLRL优化 １５
5实验与分析 １８
5.1 数据集和数据处理 １８
5.1.1 数据集 １８
5.1.2 数据处理 １９
5.2 算法评价指标 ２０
5.3 实验 ２１
5.4 结果和讨论 ２２
5.5 面部表情识别可视化 ２３
6 总结与展望 ２４
致谢 ２４
参考文献 ２５
基于标记分布学习的人脸表情情感识别研究
摘要
大多数人脸表情情感的研究都假定数据集中的表情都只具有单一的情感，符合单标记学习的研究范畴，但实际情况中的人脸表情是复杂综合的。标记分布学习(LDL)是一种新型的机器学习范式，在LDL中，标记分布类似于概率分布，可以很好应对人脸表情这种复杂标记分布的情况。先前的研究只能表述一个表情与哪些情感相关，而LDL在此基础上进一步解决了一个表情中的情感重要程度的问题。本文为基于标记分布学习的人脸表情情感识别研究，代码实现了三个基于LDL的人脸表情情感识别算法：基于局部标记相关性的标记分布识别算法 *51今日免费论文网|www.51jrft.com +Q: @351916072@ 
(LDLSCL)考虑标记的局部相关性，构造局部相关向量来反映局部样本的影响；基于局部低阶标记相关性的情感识别方法(EDLLRL)在局部相关性的基础上采用局部低阶结构来隐式利用标记相关性，从而更好地捕获复杂的标记相关性；使用标记特定特征的标记分布识别方法(LDLSF)基于标记特定特征选择和公共特征选择进一步提升了算法性能。论文介绍了对于三个算法的参数优化工作，最后在sJAFFE和sBU_3DFE两个数据集上进行大量实验，在多个评价指标下横向对比三个算法性能，并设计了可视化界面展示应用效果。
关键字：机器学习；表情识别；情感识别；标记分布学习
RESEARCH ON HUMAN FACIAL EXPRESSION RECOGNITION BASED ON LABEL DISTRIBUTION LEARNING
ABSTRACT
Most researches on facial expressions assume that the expressions in the data set have only one single emotion, which is in line with the research scope of singlelabel learning, but the actual facial expressions are complex and comprehensive. Label distribution learning (LDL) is a novel machine learning paradigm, in which the label distribution is similar to the probability distribution, which can well cope with the complex label distribution of human facial expressions. Previous research can only express which emotions an expression is related to, but LDL further solves the problem of the importance of emotions in an expression on this basis. This paper focus on the research on facial expression emotion recognition based on label distribution learning, three LDLbased facial expression emotion recognition algorithms are implemented: Label Distribution Learning by Exploiting Sample Correlations Locally(LDLSCL) considers the local correlation of labels and constructs a local correlation vector to reflect the influence of local samples; Facial Emotion Distribution Learning by Exploiting LowRank Label Correlations Locally(EDLLRL) uses local loworder structure to implicitly utilize label correlation on the basis of local correlation, so as to better capture complex label correlation; Label Distribution Learning with LabelSpecific Features(LDLSF) based on specific feature selection and common feature selection further improves algorithm performance. Paper introduces the parameter optimization work for the three algorithms，at the end of this paper, a large number of experiments were conducted on the two data sets sJAFFE and sBU_3DFE, the performance of the three algorithms was compared horizontally under multiple evaluation indicators, and a visual interface was designed to show the application effect.

原文链接：http://www.jxszl.com/jsj/wljs/606900.html