基于类别特征自注意力的人脸表情识别研究  

Research on Facial Expression Recognition Based on Category Feature Self-Attention

作　　者：李堃 宋晨光 李锐 王倩倩 Li Kun;Song Chenguang;Li Rui;Wang Qianqian(College of information&Network Engineering,Anhui Science and Technology University,Bengbu,Anhui 233000,China)

机构地区：[1]安徽科技学院信息与网络工程学院,安徽蚌埠233000

出　　处：《黑龙江工业学院学报(综合版)》2024年第12期108-113,共6页Journal of Heilongjiang University of Technology(Comprehensive Edition)

基　　金：安徽省高等学校自然科学研究项目“基于离散Gabor变换的盲源分离的关键技术研究”(项目编号:2023AH051867);安徽科技学院引进人才项目“未知网络协议的安全性测试研究”(项目编号:200192-引进人才项目XWYJ202004)。

摘　　要：人脸表情分为多种类别,部分表情类别在表达时具有一定的相似性。为了降低相似表情对分类的影响,提出了一种基于类别自注意力的表情识别方法。方法分为类别特征提取和类别特征融合两个模块。在类别特征提取模块,将CNN提取的表情图像特征通过投影,得到N份基本表情类别特征,并通过CLIP预训练模型的文本编码器和特征进行计算,获得各基本类别损失。在类别特征融合模块,将提取的基本类别特征送入Vit模型的Encoder编码器,通过自注意力机制融合以获得当前表情真实类别的最优表达。通过类别特征的提取和融合,可以降低不同类型表情特征间的相似性,消除冗余信息。方法在RafDB和CK+两个数据集上做了验证实验,通过和多个SOTA方法进行比较,证明了本文方法的有效性。Facial expressions can be divided into multiple categories,and some expression categories have certain similarities.In order to reduce the impact of similar expressions on classification,a category self-attention based expression recognition method is proposed.The method is divided into two modules:category feature extraction and category feature fusion.In the category feature extraction module,the expression image features extracted by the CNN are projected to obtain N basic expression category features,which are then calculated using the CLIP pre trained model′s text encoder and features to obtain the loss of each basic category.In the category feature fusion module,the extracted basic category features are fed into the encoder of the Vit model,and fused through a self-attention mechanism to obtain the optimal expression of the current expression′s true category.By extracting and fusing category features,it is possible to reduce the similarity between different types of facial expression features and eliminate redundant information.The method is validated on two datasets,RafDB and CK+,and its effectiveness is demonstrated by comparing it with multiple SOTA methods.

关 键 词：人脸表情识别 自注意力 特征融合 CLIP VIT 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]