Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (2): 184-192.DOI: 10.3778/j.issn.1002-8331.2008-0076

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Expression Recognition with Separable Convolution Channel Enhancement Features

LIANG Huagang, LEI Yixiong   

  1. School of Electronics and Control Engineering, Chang’an University, Xi’an 710064, China
  • Online:2022-01-15 Published:2022-01-18



  1. 长安大学 电子与控制工程学院,西安 710064

Abstract: At present, facial expression recognition network model has low recognition rate and complex parameters. To mitigate this problem, a facial expression recognition method with separable convolution channel enhancement features is proposed in this paper. First, a lightweight convolutional neural network structure is designed to extract facial features, and depthwise separable convolution is used in the convolutional layer to reduce network parameters. Then, the squeeze-and-excitation module is introduced to assign weights to the features of different channels, and different squeeze rates are used in different convolutional layers to enhance the network’s ability to extract facial expressions. Finally, the extracted features are fed into a classifier to achieve facial expression classification, and experiments are carried out and analyzed on CK+ and FER2013 dataset. The experimental results show that, compared with the existing methods, the recognition rate of the network structure proposed in this paper increases by 0.15 percentage points and 3.29 percentage points respectively on the CK+ and FER2013 dataset, and the number of network model parameters decreases by 75%. The proposed method not only reduces the network parameters, but also improves the accuracy of facial expression recognition.

Key words: facial expression, convolutional neural network, depthwise separable convolution, squeeze-and-excitation module

摘要: 针对目前人脸表情识别准确率不高、网络模型参数复杂等问题,提出一种增强可分离卷积通道特征的人脸表情识别研究方法。设计了一种轻量型卷积神经网络结构提取表情特征,在卷积层中采用深度可分离卷积减少网络参数;引入了压缩激发模块,对不同通道的特征进行权重分配,在不同的卷积层采用不同的压缩率来增强网络对人脸表情的特征提取能力;将提取到的特征送入分类器实现人脸表情分类,在CK+和FER2013数据集上进行实验并分析。实验结果表明:与现有方法相比,提出的网络结构在CK+和FER2013数据集上,识别率分别提高了0.15个百分点和3.29个百分点,且网络模型参数量降低了75%。所提方法在降低网络参数的同时,提高了表情识别准确率。

关键词: 人脸表情, 卷积神经网络, 深度可分离卷积, 压缩激发模块