基于迁移学习和注意力机制的人脸表情识别  

Facial expression recognition based on transfer learning and attention mechanism

作　　者：吕洪武[1] LYU Hongwu(School of Computer Science&Engineering,Changchun University of Technology,Changchun 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《长春工业大学学报》2022年第6期700-704,共5页Journal of Changchun University of Technology

基　　金：吉林省科技计划项目(20210201051GX)。

摘　　要：针对使用深度学习训练人脸表情识别模型过程慢,容易出现梯度爆炸和梯度消失,以及模型关注点过于分散的问题,提出一种基于迁移学习和注意力机制的人脸表情识别网络模型。文中以VGG16为主干模型,添加注意力机制和丢弃层,减少其全连接层节点数量,并使用迁移学习方法进行网络权重的初始化,在RAF-DB数据集上进行验证,达到86.05%的准确率,证明了文中算法的有效性。In order to solve the problems of slow process of training facial expression recognition model by using depth learning,gradient explosion and gradient disappearance,as well as excessive dispersion of model concerns,a network model of facial expression recognition based on transfer learning and attention mechanism is proposed.This paper takes VGG16 as the backbone model,adds attention mechanism and Dropout layer,reduces the number of full connection layer nodes,and uses migration learning method to initialize the network weight,which is verified on RAF-DB dataset,and achieves 86.05%accuracy,which proves the effectiveness of this algorithm.

关 键 词：深度学习 表情识别 迁移学习 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]