基于特征融合的注意力双线性池细粒度表情识别  

Fine-grained Expression Recognition of Attention Bilinear Pooling Based on Feature Fusion

作　　者：刘力源 张利锋[1] 柳婵娟[1] 贾世祥[1] LIU Liyuan;ZHANG Lifeng;LIU Chanjuan;JIA Shixiang(School of Information and Electrical Engineering,Ludong University,Yantai 264039,China)

机构地区：[1]鲁东大学信息与电气工程学院,山东烟台264039

出　　处：《鲁东大学学报（自然科学版）》2020年第2期130-136,共7页Journal of Ludong University:Natural Science Edition

基　　金：山东省自然科学基金(ZR2016FM23);山东省重点研究与发展计划(2017GHY215009)。

摘　　要：面部表情的细微差异可能传达出截然不同的情绪,使得表情识别成为一项颇具挑战的任务.本文提出一种基于特征融合的注意力双线性池模型FFABP,用于捕获、融合二阶局部特征和一阶全局特征,实现细粒度特征与粗粒度特征的融合.该模型在双线性池模型的基础上引入注意力机制,利用双线性池模型的高维空间表示捕获表情之间的细微局部差别,借助注意力机制突出特征图中重要空间位置的作用;并使用自注意力模型来学习不同区域特征之间的依赖关系,从而获得图像的全局几何特征.实验表明,与现有的其它模型相比,本文模型在数据集FER2013和CK+上能够获得较好的性能表现和鲁棒性.The subtle differences in facial expressions can convey very different emotions,making facial recognition a challenging task.In this paper,a feature fusion attention bilinear pooling model(FFABP)has been proposed to capture and fuse second-order local features and first-order global features,which realizes the combining of the coarse-grained and fine-grained features.Attention is used by bilinear pooling model.The high-dimensional space representation of the bilinear pooling model is used to capture the subtle local differences between expressions,and the attention is used to highlight the role of important spatial positions in the feature maps.Self-attention is used in FFABP to learn the relationship between different regional features,which can obtain the global geometric features of images.Experiments show that the model can obtain competitive performance and robustness on FER2013 and CK+datasets compared with other existing models.

关 键 词：特征融合 注意力机制 细粒度表情识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]