| 摘 要: 为了提高训练速度和人脸表情识别效果,提出一种基于局部二值模式(Local Binary Pattern,LBP)和注意力机制的改进视觉几何群网络(Visual Geometry Group Network,VGG网络)的人脸表情识别方法。首先,通过LBP获取数据集的纹理特征。其次,利用全局平均池化层代替全连接层,并在基准模型卷积层后和全局平均池化层前引入注意力模块,创建新网络模型NEW-VGG;通过对NEW-VGG做消融实验,验证模型改进的正确性。最后,融合LBP+NEW-VGG模型对CK+和Fer2013两种数据集进行10倍交叉验证,取得了97.98%和76.75%的识别率。实验结果表明,该方法不仅能加快网络训练迭代速度,增强人脸表情识别效果,还具有较强的鲁棒性。 |
| 关键词: 面部表情识别;局部二值模式;注意力机制 |
|
中图分类号: TP391
文献标识码: A
|
|
| Improved VGG Network for Facial Expression Recognition Method Based on LBP and Attention Mechanism |
|
ZHANG Zhonghua1, YANG Huijiong2
|
(1.School of Computer Science and Technology, TaiYuan Normal University, Jinzhong 030619, China;
2.Taiyuan Institute of Technology, Taiyuan 030008, China)
1940219034@qq.com; yanghj@tit.edu.cn
|
| Abstract: In order to improve training speed and facial expression recognition performance, this paper proposes an improved VGG network (Visual Geometry Group Network) based on LBP (Local Binary Pattern) and attention mechanism for facial expression recognition. Firstly, the texture features of the dataset are obtained through LBP. Secondly, the fully connected layer is replaced with the global average pooling layer and an attention module is introduced after the benchmark model convolution layer and before the global average pooling layer, so that a new network model NEW-VGG is created. Ablation experiment on NEW-VGG verifies the correctness of the model improvement. Finally, the LBP+NEW-VGG fusion model is subjected to 10-fold cross validation on the CK+ and Fer2013 datasets, achieving recognition rates of 97.98% and 76.75% . The experimental results show that the proposed method not only accelerates the iteration speed of network training and enhances the facial expression recognition effect, but also has strong robustness. |
| Keywords: facial expression recognition; Local Binary Pattern; attention mechanism |
