一种基于注意力机制的轻量级面部情绪识别方法  

A Lightweight Facial Emotion Recognition Method Based on Attention Mechanism

作　　者：袁佩瑶 宋斌 毛岳恒[1,2] 张志勇 YUAN Pei-yao;Song Bin;MAO Yue-heng;ZHANG Zhi-yong(Information Engineering College,Henan University of Science and Technology;Henan International Joint Laboratory of Cy-berspace Security Applications,Henan University of Science and Technology,Luoyang 471023,China)

机构地区：[1]河南科技大学信息工程学院 [2]河南科技大学河南省网络空间安全应用国际联合实验室,河南洛阳471023

出　　处：《软件导刊》2022年第12期7-13,共7页Software Guide

基　　金：国家自然科学基金项目(61972133);河南省中原千人计划中原科技创新领军人才项目(204200510021);河南省重点科技攻关项目(202102210162,212102210383)。

摘　　要：针对传统人脸表情识别研究中存在的模型训练、应用落地以及小样本学习困难等关键技术问题和挑战,提出一种基于注意力机制的轻量级面部情绪识别方法。该方法对Xception网络进行改进,增加了注意力模块作为辅助块,分别通过通道注意力模块和空间注意力模块使模型注意力集中在有意义的图像输入和表情强度更大的区域中。利用Switchable Normalization(SN)归一化,为处在神经网络不同位置、不同深度的归一化层选择不同操作,赋予有利于网络层的归一化方式更大权重。通过在FER2013数据集上与其他经典模型进行实验对比,该方法准确率可达到86%,其中在生气、开心、惊讶和自然的识别率较高,均能达到90%以上。另外针对不受约束的人脸表情场景,该模型对传统卷积神经网络进行了改进,弥补了传统卷积神经网络参数量大、训练难、落地难、梯度消失以及梯度爆炸等缺陷。Aiming at the key technical problems and challenges existing in traditional facial expression recognition research, such as model training, application implementation and small sample learning difficulties, propose a lightweight facial emotion recognition method based on attention mechanism. In this method, the Xception network is improved by adding an attention module as an auxiliary block, and the model’s attention is focused on the meaningful image input and the region with greater expression intensity through the channel attention module and the spatial attention module respectively. The Switchable Normalization(SN) is used to select different operations for normalization layers at different locations and depths in the neural network, giving greater weight to normalization approaches that benefit the network layer. Compared with other classical models on FER2013 dataset, the accuracy of this method can reach 86%, and the recognition rate of angry, happy, surprised and natural is higher, which can reach more than 90%. In addition, the model improves the traditional convolutional neural network for unconstrained facial expression scenes, and makes up for the defects of traditional convolutional neural network such as large number of parameters, difficulty in training, difficulty in landing, gradient disappearance and gradient explosion.

关 键 词：面部情绪识别 深度学习 注意力模型 轻量级网络 辅助块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]