摘要
针对轻量级面部表情识别算法泛化能力的不足,提出了一种结合多特征融合和注意力机制的表情识别方法。使用局部二值模式(Local Binary Pattern,LBP)算子减少面部图像中无关信息的干扰,双分支神经网络提取原始人脸图像和LBP图像的特征,融合两个网络提取的中高层特征,并通过注意力机制加强重要特征,在保持较少参数量的同时生成大量的有效特征信息提高算法的识别效果。实验结果表明,该方法在Fer2013和CK+数据集上的识别率分别为70.21%和95.59%,有效地提高了轻量级表情识别算法的性能。
For the insufficient generalization ability of lightweight facial expression recognition algorithms,a facial expression recognition method based on multi-feature fusion and attention mechanism is proposed.LBP is used to reduce the interference of meaningless information of original face image,the dual-branch neural network extracts the features of original face image and LBP image respectively,merges the middle-level and high-level feature information extracted by these two networks and strengthens important features through attention mechanism for facial expression recognition,which can generate a large amount of discriminative features by multi-feature fusion to improve the recognition performance with a small number of parameters.The experimental results show that the recognition rate of this method has reached 70.21%on Fer2013 dataset and 95.59%on CK+dataset respectively,which effectively improves the performance of lightweight expression recognition algorithm.
作者
赖东升
冯开平
罗立宏
Lai Dong-sheng;Feng Kai-ping;Luo Li-hong(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China;School of Art and Design,Guangdong University of Technology,Guangzhou 510090,China)
出处
《广东工业大学学报》
CAS
2023年第3期10-16,共7页
Journal of Guangdong University of Technology
基金
教育部人文社科资助项目(20YJAZH073)。
关键词
人脸表情识别
轻量级网络
局部二值模式
多特征融合
注意力机制
facial expression recognition
lightweight network
local binary pattern
multi-feature fusion
attention mechanism