Abstract:

Facial expression recognition has widely application prospects in human-computer interaction, such as clinical medicine, safe driving. To solve the problem of low accuracy due to the inability of the LSTM network to effectively extract correlative information prior and post frames, the framework is proposed based on Bi-LSTM, while the theory of identity mapping residual is used to effectively prevent the gradient disappearance caused by network depth. Because the Inception-V3 network has too many parameters and easy over-fitting, proposed to add two Reduction modules to reduce the parameters, and then get the more generalized model of Inception-w. Finally, the designed model has been evaluated on two datasets, CK+ and Oulu-CASIA, and compared with state-of-the-art methods. The highest accuracy reaches 99.6%, which indicates that the proposed method has better recognition accuracy within a certain range.

Key words: facial expression recognition, Inception-w model, Res-Bi-LSTM, spatio-temporal features

摘要：

人脸表情识别在人机交互、临床医学、安全驾驶等方面有着广泛的应用前景。针对传统LSTM网络只能根据单向传播信息学习表情时序特征的局限，提出在双向传播的Bi-LSTM网络基础上，采用恒等映射残差理论有效防止易受网络深度引起的梯度消失问题。又因为提取空间特征的Inception-V3网络存在参数过多，容易过拟合等问题，提出添加两个Reduction模块减少参数，进而得到泛化性更好的Inception-w模型。最后对设计的模型在CK+和Oulu-CASIA两个数据集上进行实验，并与现有方法进行对比。实验可得最高识别率为99.6%，表明该方法在一定范围内具有较好的识别准确率。

关键词: 表情识别, Inception-w模型, Res-Bi-LSTM, 时空特征