动态多阶段强波动型表情识别模型  

Dynamic multi-stage model for strong fluctuating facial expression recognition

作　　者：欧阳勇[1] 陈凌钰 曾雅文 万俊 王春枝[1] OUYANG Yong;CHEN Ling-yu;ZENG Ya-wen;WAN Jun;WANG Chun-zhi(School of Computer Science,Hubei University of Technology,Wuhan 430068,China;College of Computer Science and Software Engineering,Shenzhen University,Shenzhen 518060,China)

机构地区：[1]湖北工业大学计算机学院,湖北武汉430068 [2]深圳大学计算机与软件学院,广东深圳518060

出　　处：《计算机工程与设计》2021年第10期2970-2978,共9页Computer Engineering and Design

基　　金：国家自然科学基金面上基金项目(61772180);全国大学生创新创业训练计划基金项目(201910500010)。

摘　　要：考虑到问卷调查顾客的满意度评价时存在样本获取困难、主观误差等问题,利用表情识别技术分析视频中顾客的情感表达,但实际应用存在人物表情变化波动过大的强波动型表情识别问题。为更好地挖掘强波动型表情变化的信息,提出一种动态多阶段强波动型表情识别模型(LNsCo),该模型包括长期情绪波动状态生成器(LEF)、近短期表情特征生成器(NsSE)、平行共同注意力网络(co-attention network)。将图像序列预处理后,分别送到LEF和NsSE提取表情的隐藏特征,利用共同注意网络生成二者相互依赖的表征,用分类器进行表情分类。在公共数据集和真实应用场景下的结果表明,所述方法具有良好的性能。In view of sample acquisition difficulty and subjective fault by obtaining customers satisfaction evaluation in the form of questionnaire,the expression recognition technology is used to analyze the emotional expression of customers in the video,but there is a strong fluctuation facial expression recognition difficulty in the actual application.To better mine the information of strong fluctuating expression changes,a dynamic multi-stage model for strong fluctuating facial expression recognition(LNsCo)was proposed,which mainly included a long-term emotional fluctuation states generator(LEF),a near-short-term expression features generator(NsSE),and a parallel co-attention network(co-attention network).After image sequences were preprocessed,they were sent to LEF and NsSE to extract the hidden features of the expression images,and co-attention network was designed to generate the interdependent representation of the two.Classifier was used to classify the expression.The results under public dataset and real application scenarios show that the proposed method has good performances.

关 键 词：强波动型表情识别 长期情绪波动状态生成器 近短期表情特征生成器 平行共同注意力网络 双向长短期记忆网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]