基于数据融合的表情识别

发布时间：2018-08-05 16:09

【摘要】：为了解决表情识别中单一数据所包含人脸表情信息不全面的问题,融合了图像与标记点数据;针对传统模式识别方法中手动提取特征的复杂性,采用神经网络框架,从而实现了特征的自动提取。本文算法以人脸表情的图像与标记点数据为基础,以神经网络为框架,采用稀疏自动编码器对网络进行预训练,实现了网络的稀疏连接,另外,在网络权值更新过程中结合了结构化正则项(structured regularization),限制了不同数据与隐层神经元的连接。实验表明:图像与标记点数据的融合更全面地表达了人脸表情信息;稀疏自动编码器和结构化正则项的运用能更有效地提取关键特征,并使神经网络自动分析不同输入数据在表情识别中所起到的作用强弱。
[Abstract]:In order to solve the problem of incomplete facial expression information contained by single data in facial expression recognition, the image and marker data are fused. In view of the complexity of manual feature extraction in traditional pattern recognition methods, neural network framework is adopted. Thus the automatic feature extraction is realized. This algorithm is based on facial expression image and marker point data, takes neural network as the frame, uses sparse automatic encoder to pre-train the network, realizes the sparse connection of the network, in addition, In the process of network weight updating, the structured regular term (structured regularization), is used to restrict the connection between different data and hidden layer neurons. The experimental results show that the fusion of image and marker data can more fully express facial expression information, and the application of sparse automatic encoder and structured canonical items can extract the key features more effectively. The neural network can automatically analyze the effect of different input data in facial expression recognition.
【作者单位】：山东大学控制科学与工程学院;
【基金】：国家自然科学基金资助项目(61573222;61203253;61233014) 山东省科技重大专项项目(新兴产业)(2015ZDXX0801A02) 山东大学基本科研业务费专项资金资助项目(2016JC014) 江苏省三维打印装备与制造重点实验室开放课题资助项目(3DL201502)
【分类号】：TP391.41

【相似文献】