摘要
面部表情识别已经广泛运用于人机交互、教育、自动驾驶等各种领域,针对现有表情识别方法网络结构复杂、参数多、泛化能力不足、识别率低等问题,且人脸表情在自然环境下易受到光照、姿态和局部遮挡等环境因素的影响,提出了一种多尺度残差注意力网络.以ResNet-18网络为基础,提出了一种新的多尺度残差注意力模块,通过提取不同尺度特征,增加特征多样性.引入CBAM注意力机制获取表情图像重点特征信息,有利于提升遮挡表情的识别.构建特征残差融合块,将浅层特征与深层特征进行残差融合,有利于获取丰富的人脸表情图像整体特征.实验结果表明,本文方法在CK+、JAFFE和Oulu-CASIA 3个公开表情数据集上分别达到了99.49%、98.53%和97.08%的准确率,与一些现有方法相比,本文方法表情识别率更高,证明了该方法可用于人脸表情识别.
Facial expression recognition has been widely used in human-computer interaction,education,automatic driving and other fields.Aiming at the problems of complex network structure,many parameters,insufficient generalization ability and low recognition rate of existing facial expression recognition methods,and the influence of environmental factors such as illumination,posture and local occlusion on facial expressions in natural environment,a multi-scale residual attention network is proposed.Based on ResNet-18 network,a new multi-scale residual attention module is proposed to increase feature diversity by extracting different scale features.CBAM attention mechanism is introduced to obtain the key feature information of facial expression images,which is beneficial to improve the recognition of occluded facial expressions.A feature residual fusion block is constructed to fuse shallow features and deep features,which is beneficial to obtain rich overall features of facial expression images.Experimental results show that the proposed method achieves 99.49%,98.53%and 97.08%accuracy on CK+,JAFFE and Oulu-CASIA public expression datasets,respectively.Compared with some existing methods,the proposed method has higher expression recognition rate,which proves that the method can be used for facial expression recognition.
作者
袁德荣
张勇
唐颖军
李波燕
谢宝来
YUAN Derong;ZHANG Yong;TANG Yingjun;LI Boyan;XIE Baolai(School of Software and Internet of Things Engineering,Jiangxi University of Finance and Economics,Nanchang 330013,China;School of Statistics,Jiangxi University of Finance and Economics,Nanchang 330013,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2024年第1期30-36,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61762043)资助
江西省教育厅自然科学研究项目(GJJ210507,GJJ190249,GJJ160425)资助.
关键词
人脸表情识别
多尺度特征
CBAM注意力机制
遮挡人脸识别
facial expression recognition
multi-scale features
CBAM attention mechanism
occlusion face recognition
