Abstract:

Facial expression recognition is one of the hotspots in the field of computer vision. Aiming at the problems that the extracted high-dimensional features recognition rate of facial expression images is poor and the generalization is almost unsatisfactory, combined with the advantages of convolutional neural network to learn local features, this paper proposes a method of multi-channel, image segmentation and optimization of Gabor convolution kernel parameters to realize image convolution, and analyzes the extracted expression adaptive features. Automatic encoder theory is used to realize high dimensional feature dimension reduction and multi-channel feature fusion. Because the traditional SVM（Support Vector Machine） is not sensitive to multi-classification tasks, the genetic algorithm is used to optimize the maximum interval classification surface and obtain fitter classifier parameters. Finally, the designed GaAeS-net（Gabor Autoencoder Support Vector Machine Convolution Network） is tested on CK+, JAFFE, FER2013, CHD2018 and other databases, and compared with the existing models. The highest recognition rate reaches 99.34% by GaAeS-net, which proves that the model has a good recognition rate and generalization.

Key words: expression recognition, Gabor kernel, convolutional neural network, automatic encoder, genetic algorithm, Support Vector Machine（SVM）, parameter optimization

摘要：

人脸表情识别是计算机视觉领域研究的热点之一。针对传统Gabor网络提取表情图像高维特征识别率不高、泛化性不强的问题，结合卷积神经网络学习局部特征的优点，提出多通道、图像分块、优化Gabor卷积核参数的方法实现表情图像卷积，对提取的表情自适应特征进行分析，首先进行通道内降维，然后采用自动编码器理论解决高维特征降维和多通道特征融合的问题。因为传统支持向量机（Support Vector Machine，SVM）对多分类问题不敏感，所以采用遗传算法优化出最大间隔分类面，进而得到适应度较高的分类器参数。对设计的GaAeS-net（Gabor Autoencoder Support Vector Machine Convolution Network）网络分别在CK+、JAFFE、FER2013、CHD2018等数据库上进行实验，并与现有模型进行对比，最高识别率可达到99.34%，从而证明GaAeS-net模型具备良好的识别率和泛化性。

关键词: 表情识别, Gabor核, 卷积神经网络, 自动编码器, 遗传算法, 支持向量机（SVM）, 参数优化