摘要
针对基于卷积神经网络(CNN)的人脸表情识别方法对图像细节特征提取不充分,模型泛化能力较差,关键特征信息利用率低的问题,提出一种多尺度感受野增强的人脸表情识别网络(MRENet)。首先,设计融合空洞卷积的多尺度特征提取模块(AMF Block),该模块在不增加参数量的同时,融合不同感受野特征图信息,提高网络的准确率与鲁棒性;然后,融合通道注意力和空间注意力的双注意力机制(DAM),能够充分提取通道域和空间域特征,增强对关键特征的感知能力与提取能力,提升网络识别精度;最后,采用SoftMax分类器实现对人脸表情的准确识别。本文方法在数据集FER2013、CK+和JAFFE上识别准确率分别达到74.76%、98.49%和98.42%,证明了该方法的有效性。
Aiming at the problems that the facial expression recognition method based on convolutional neural network(CNN)does not fully extract the detailed features of the image and has poor model generalization ability,and has low utilization rate of key feature information,a multi-scale receptive field enhanced facial expression recognition network(MRENet)is proposed.Firstly,a multi-scale feature block with fusion atrous convolution(AMF Block)is designed.This module integrates feature map information from different receptive fields without increasing the parameter quantity,thereby enhancing the accuracy and robustness of the network.Secondly,a double attention mechanism(DAM)that fuses channel attention and spatial attention can fully extract channel domain and spatial domain features,enhance the perception and extraction capabilities of key features,and improve network recognition precision.Finally,the SoftMax classifier is used to achieve accurate recognition of facial expressions.The recognition accuracy of this method on the datasets FER2013,CK+and JAFFE reach 74.76%,98.49%and 98.42%,respectively,which proves the effectiveness of the method.
作者
袁姮
常峻溪
YUAN Heng;CHANG Junxi(Software College,Liaoning Technical University,Huludao 125100,China)
出处
《传感器与微系统》
北大核心
2025年第6期51-55,共5页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(61172144)
辽宁省自然科学基金资助项目(20170540426)
辽宁省教育厅重点基金资助项目(LJYL049)。
关键词
人脸表情识别
空洞卷积
注意力机制
多尺度特征
facial expression recognition
atrous convolution
attention mechanism
multi-scale feature
作者简介
袁姮(1988-),女,副教授,主要研究领域为模式识别与人工智能、图像与视觉信息计算;通讯作者:常峻溪(1998-),男,硕士研究生,研究方向为模式识别与人工智能、图像与视觉信息计算。
