| 摘要: |
| 针对当前人脸表情识别算法识别精度不高、网络鲁棒性差的缺点，设计了一种改进型Dense HRNet特征提取网络，使用稠密连接机制强化了HRNet中浅层特征与深层特征间的传递和融合方式。同时，提出了一种基于基尼指数动态加权决策算法，根据每一卷积神经网络(Convolutional Neural Network,CNN)支路分类的确定性，为各支路输出动态地赋予权重，提高多路CNN支路融合决策的准确性，解决了由于单路CNN分类不确定性引起的偶然误差。在FER2013数据集和CK+数据集上进行实验，所提方法分类准确率分别达到73.36%和97.59%。 |
| 关键词: 人脸表情识别；Dense HRNet；基尼指数 加权决策 |
| DOI： |
|
| 基金项目: |
|
| An expression recognition algorithm based on improved Dense-HRNet and Gini-index dynamic weighted decision |
| LAN Zhengjie,WANG Lie,HUANG Ying |
| (School of Computer and Electrics Information,Guangxi University) |
| Abstract: |
| For the shortcomings of the current facial expression recognition algorithms,such as low recognition accuracy and poor network robustness,a feature extraction network based on the improved Dense HRNet is designed,and the dense connection mechanism is used to enhance the transmission and fusion of features with different resolutions in HDNet.At the same time,a dynamic weighted decision classification algorithm based on the Gini index is proposed.According to the confidence degree of the classification probability,the fusion weight is automatically assigned to the output probability of each convolutional neural network(CNN) branch,in this way the network solves the instability of the single channel CNN classification.Experiments on the FER2013 dataset and CK+ dataset show that the proposed method achieves the classification accuracy of 73.36% and 97.59%,respectively. |
| Key words: facial expression recognition Dense HRNet Gini index weighted decision |
