基于动态贝叶斯网络的头姿无关自发表情识别研究

项目名称： 基于动态贝叶斯网络的头姿无关自发表情识别研究

项目编号： No.61463034

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 何俊

作者单位： 南昌大学

项目金额： 46万元

中文摘要： 表情识别研究热点正从摆拍表情转移到自发表情，但连续头姿运动阻碍了表情识别在手机等智能终端领域的应用。本课题围绕自发表情识别技术中头姿无关的表情特征提取、分类决策模型这两个关键技术展开研究。首先，根据高斯过程回归算法在小样本下对未知数据的泛化能力，研究利用有限离散头姿下的表情训练数据推测连续头姿下的表情特征问题的新方法，特别是对高斯过程回归算法模型中存在的计算量大、噪声必须服从高斯分布以及多类输出等问题开展进一步研究；其次，鉴于目前大多研究对头姿估计、脸部肌肉分别建模，忽视了头姿也是表情特征之一这个本质，拟运用动态贝叶斯网络对连续头姿和脸部肌肉联合建模，解决头姿与脸部肌肉的非线性耦合关系，为自发表情识别提供科学的理论依据，同时为变量的非线性时空解耦、融合与决策研究开辟一个新的领域，必将具有广阔的应用前景。

中文关键词： 自发表情识别；动态贝叶斯网络；头姿无关；高斯过程回归

英文摘要： Research on facial expression recognition has steadily been moving from posed frontal expressions to spontaneous expressions. but continuous head pose movement is key problem for application of spontaneous expression recognition such as smart phones. This project is mainly about two key problem in spontaneous expression recognition: expression feature extraction under continuous head posture and classification decision model. Firstly, according to the generalization ability of Gauss process regression algorithm in small samples of unknown data. Study a new method to get expression feature under continuous head posture using the expression training data of finite discrete head posture. Especially for the problem of complex calculation, noise must obey Gauss distribution and multi class output existing in Gauss process regression algorithm, which should be further studied. Secondly, considering the current most research build model on pose estimation, facial muscles respectively, Ignoring the fact that the head pose should belong to facial expression feature, just like facial muscle. We plan to model head pose and facial muscles based on Dynamic Bayesian Network jointly, solve the nonlinear coupling existing between the head pose and facial muscles. provide scientific theoretical basis for the spontaneous expression recognition. meanwhile, open up a new field for the study of nonlinear decoupling in time and space, fusion, decision between variables, which should be have broad application prospects.

英文关键词： spontaneous expression recognition;Dynamic Bayesian Network;head pose independent;Gaussian Process Regression