摘要
针对人脸表情呈现方式多样化以及人脸表情识别易受光照、姿势、遮挡等非线性因素影响的问题,提出了一种深度多尺度融合注意力残差网络(deep multi-scale fusion attention residual network,DMFA-ResNet)。该模型基于ResNet-50残差网络,设计了新的注意力残差模块,由7个具有三条支路的注意残差学习单元构成,能够对输入图像进行并行多卷积操作,以获得多尺度特征,同时引入注意力机制,突出重点局部区域,有利于遮挡图像的特征学习。通过在注意力残差模块之间增加过渡层以去除冗余信息,简化网络复杂度,在保证感受野的情况下减少计算量,实现网络抗过拟合效果。在3组数据集上的实验结果表明,本文提出的算法均优于对比的其他先进方法。
This paper proposes a deep multiscale fusion attention residual network based on the ResNet-50 model to solve the problems of the diversification of facial expression presentation and the susceptibility of facial expression re-cognition to nonlinear factors,such as illumination,posture,and occlusion.A novel attention residual module consist-ing of seven attention residual learning units with three branches is designed to perform multiple convolution operations on the input image in parallel and obtain multiscale features.To highlight important local areas,the attention mechan-ism is introduced simultaneously,which is conducive to the feature learning of the occluded images.Furthermore,a novel transition layer is added between the attention residual modules to remove redundant information,simplify the network complexity,reduce the amount of calculation while ensuring the receptive field,and realize the anti-overfitting effect of the network.Experimental results on three datasets demonstrate that the proposed algorithm is superior to other advanced methods.
作者
高涛
杨朝晨
陈婷
邵倩
雷涛
GAO Tao;YANG Zhaochen;CHEN Ting;SHAO Qian;LEI Tao(School of Information Engineering,Chang’an University,Xi’an 710000,China;School of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an 710021,China)
出处
《智能系统学报》
CSCD
北大核心
2022年第2期393-401,共9页
CAAI Transactions on Intelligent Systems
基金
国家重点研发计划项目(2019YFE0108300)
国家自然科学基金项目(62001058)
陕西省重点研发计划项目(2019GY-039)
长安大学中央高校基本科研业务费专项资金项目(300102241201).
关键词
人脸表情识别
残差网络
多尺度特征
注意力机制
遮挡人脸
卷积神经网络
特征融合
深度学习
facial expression recognition
residual network
multiscale features
attention mechanism
occlusion of hu-man faces
convolution neural network
feature fusion
deep learning
