基于贝叶斯网和迁移学习的人脸表情识别方法  

Facial expression recognition method based on Bayesian network and transfer learning

作　　者：郭文强[1] 冯春石 赵艳[2] 侯勇严[2] 徐成 李惟[3] GUO Wen-qiang;FENG Chun-shi;ZHAO Yan;HOU Yong-yan;XU Cheng;LI Wei(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi′an 710021,China;School of Electrical and Control Engineering,Shaanxi University of Science&Technology,Xi′an 710021,China;School of Electronics and Control Engineering,Chang′an University,Xi′an 710054,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021 [2]陕西科技大学电气与控制工程学院,陕西西安710021 [3]长安大学电子与控制工程学院,陕西西安710054

出　　处：《陕西科技大学学报》2022年第6期163-169,共7页Journal of Shaanxi University of Science & Technology

基　　金：陕西省科技厅重点研发计划项目(2020SF-286);陕西省交通运输厅科研计划项目(21-07X);陕西省西安市科技计划项目(2019216514GXRC001CG002GXYD1.1)。

摘　　要：针对人脸表情识别建模过程中获取的样本数据稀少、模型确定困难等问题,提出了一种基于贝叶斯网和迁移学习的人脸表情识别方法.首先提取人脸表情图像特征构成面部动作单元(AU)特征样本集,然后通过分析面部表情与AU关系建立人脸表情识别贝叶斯网络(BN)结构;其次求得源域BN参数和目标域BN初始参数,引入迁移机制和平衡因子自适应地进行目标网络的参数学习,建立人脸表情识别BN模型;最后利用BN推理算法实现人脸表情识别.实验结果表明,在小数据集条件下,该方法相比支持向量机、Adaboost和卷积神经网络等识别方法具有更高的识别精度.Aiming at the problems of sparse sample data and difficult model determination in the face expression recognition modeling process,a face expression recognition method based on Bayesian network and transfer learning is proposed.Firstly,the facial expression image features are extracted to form the facial Action Unit(AU)feature sample set,and then the facial expression recognition Bayesian network(BN)structure is established by analyzing the relationship between facial expressions and AUs;secondly,the source domain BN parameters and target domain BN initial parameters are obtained,the introduction of the migration mechanism and balance factor to adaptively learn the parameters of the target network,and the BN model of facial expression recognition is established;finally,the BN inference algorithm is used to realize facial expression recognition.The experimental results show that under the condition of small data sets,the method has higher recognition accuracy than the recognition methods such as support vector machine,Adaboost and convolutional neural network.

关 键 词：表情识别 迁移学习 贝叶斯网络 AU特征 BN参数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]