DFNet:融合多尺度特征与自注意力的表情识别算法  

DFNet:Expression recognition algorithm based on fusion of multi-scale features and self attention

作　　者：林帅男 张伟[1] 胡敏 赵瑞[1] LIN Shuainan;ZHANG Wei;HU Min;ZHAO Rui(College of Mathematics and Computer,Jilin Normal University,Siping 136000,Jilin,China)

机构地区：[1]吉林师范大学数学与计算机学院,吉林四平136000

出　　处：《智能计算机与应用》2024年第6期64-70,共7页Intelligent Computer and Applications

基　　金：吉林省科技厅科研项目(20230101243JC)。

摘　　要：为解决人脸表情识别时存在的特征表达能力不足以及识别率不高的问题,提出了一种新的融合多尺度特征与自注意力的表情识别算法-DFNet。进行多尺度特征融合时,通过采用空洞卷积以及通道降维的形式,在扩大感受野的同时,获得多尺度信息;提出一种快自注意力机制,改进了传统的Transformer block,提升了模型的特征提取能力,进一步提高了模型性能。实验结果表明,所提方法在RAF-DB和KDEF表情数据集分别取得了89.31%和89.05%的准确率,证明了所提网络具有较强的泛化性。To address the problems of large model size,high computational complexity,and low recognition rate in facial expression recognition,a new expression recognition algorithm based on fusion of multi-scale features and self attention mechanism-DFNet is proposed.When performing multi-scale feature fusion,by using dilated convolution and channel dimensionality reduction,multi-scale information is obtained while expanding the receptive field.A fast self attention mechanism is proposed that improves the traditional Transformer block and reduces the number of parameters,therefore improves the performance of the model.The experimental results show that the proposed method achieves accuracy of 89.31%and 89.05%on the RAF-DB and KDEF expression datasets,respectively,proving that the proposed network has strong generalization ability.

关 键 词：人脸表情识别 残差网络 自注意力机制 空洞卷积 多尺度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]