| 摘要: |
| 为了解决传统卷积神经网络（Convolutional Neural Network,CNN）对于人脸微表情识别的泛化能力差的问题，提出了一种改进的Inception结构与残差结构结合的卷积神经网络方法。首先在改进的Inception结构的基础上将输入特征直接映射到输出结果中构成残差结构，并针对表情局部特征复杂模糊等不足采用多层池化的方式进行优化，实现端到端的人脸表情识别。为防止训练数据量少、数据分布不均匀，采用了数据增强技术从原始数据集中生成更多的训练样本。新模型在Fer2013数据集上进行测试，准确率达到71.26%，表明该方法相比于传统的卷积神经网络具有更高的准确率，训练模型更加稳定。 |
| 关键词: 表情识别 卷积神经网络 Inception结构 多层池化 |
| DOI： |
|
| 基金项目:国家自然科学青年基金项目（11701060）；重庆市科委面上项目（cstc2020jcyj-msxm1397）；重庆市教委项目（KJQN202000601）；重庆师范大学开放课题项目（CSSXKFKTM202007） |
|
| Application of convolutional neural network with multi-layer pooling optimization in facial expression recognition |
| CHEN Jiachang,XIAO Sa,ZHOU Weisong |
| (School of Science,Chongqing University of Posts and Telecommunications,Chongqing 400065,China) |
| Abstract: |
| In order to overcome the poor generalization problem of facial expression recognition with traditional convolutional neural network(CNN),a new CNN method combining Inception structure with residual structure is proposed.Based on the improved Inception structure,the input features are directly mapped to the output results to form a residual structure.In view of the complex and fuzzy local features of facial expressions and other shortcomings,the multi-layer pooling method is used for optimization to realize end-to-end facial expression recognition.Owing to the lack of training data,or the unevenness of information,the over-fitting often occurs.For the purpose of avoiding this condition,data enhancement technique has been applied.Experiments on Fer2013 dataset show that the accuracy rate of this method is 71.26%,and better recognition efficiency and robustness is achieved compared with other CNN methods. |
| Key words: facial expression recognition convolutional neural network Inception structure multi-layer pooling |
