基于改进ResNet和损失函数的表情识别  

Expression Recognition Based on Improved ResNet and Loss Function

作　　者：谢银成 黎曦[1] 李天 李聪聪 XIE Yin-cheng;LI Xi;LI Tian;LI Cong-cong(Electrical Information System,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]武汉工程大学电气信息学院,武汉430205

出　　处：《自动化与仪表》2022年第4期64-69,共6页Automation & Instrumentation

基　　金：武汉工程大学研究生教育创新基金项目(CX2020066)。

摘　　要：在人脸表情识别的研究中,运用深度学习方法训练网络,需要大量标注准确的数据,但由于人脸表情的复杂性,数据集存在类别数量不均衡问题,并且由于人脸部的特征比较复杂,难以精确提取对表情识别有用的特征,对人脸表情识别的研究是一个很大的挑战。针对这些问题,该文提出一种基于改进ResNet网络和损失函数的人脸表情识别方法。首先,对ResNet网络进行改进,在网络中嵌入CBAM注意力机制模块,提升模型的关键特征提取能力;然后,针对数据集类别数量不均衡影响模型识别性能的问题,采用数据集增强和加权损失函数的方法对模型进行优化。通过实验验证,该文构建的网络模型在CK+和Fer2013表情数据集上识别效果得到提升,优于部分同类型的方法。In facial expression recognition research,utilizing the method of deep learning training network,need a lot of annotation accurate data,but because of the complexity of the facial expression,problems category number unbalanced data sets,and because the characteristics of face is more complex,it is difficult to accurately extract useful features for facial expression recognition,the study of facial expression recognition is a big challenge.To solve these problems,a facial expression recognition method based on improved ResNet network and loss function is proposed.Firstly,the ResNet network is improved,and the CBAM attention mechanism module is embedded in the network to improve the key feature extraction ability of the model.Then,aiming at the problem that the model recognition performance is affected by the unbalanced number of data set categories,data set enhancement and weighted loss function are used to optimize the model.Through experimental verification,the network model constructed in this paper has improved its recognition effect on CK+and Fer2013 expression data sets,and is superior to some methods of the same type.

关 键 词：表情识别 ResNet网络 注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]