基于角度距离损失与小尺度核网络的表情识别

针对人脸表情类内差异大、类间相似度高导致识别率低的问题，提出了一种基于角度距离损失与小尺度核网络的表情识别方法。网络基于3×3卷积核，在网络中加入融合空间金字塔注意力的点积残差块，引入Dropblock正则化，并提出了低层特征掩膜化。该模型低层特征具备高层特征的语义信息，而且参数量较少，结构简单有效。训练时，使用提出的基于角度距离损失函数监督神经网络学习，提高了网络的类间特征分离和类内特征聚类的特征判别能力。实验结果表明，该方法在CK+和FER2013数据集上识别准确率分别达到了97.88%和72.81%，具有较强竞争力。此外，消融实验表明所提出的改进方法可提高表情识别率，进一步验证了其有效性。

Facial expression recognition based on angular distance loss function and convolutional neural network

In order to solve the problem of low recognition rate caused by large intra-class differences and high similarity between classes,this paper proposes an expression recognition method based on angular distance loss and small-scale kernel network.The network is based on 3×3 convolution kernel,and a dot product residual block which fuses the attention of spatial pyramid is added to the network.Dropblock regularization is introduced,and low-level feature masking is proposed.The low-level features of the model have semantic information of high-level features with few parameters and simple and effective structure.During training,the proposed loss function based on angle distance is used to supervise the learning of neural network,which improves the feature discrimination ability of feature separation between classes and feature clustering within classes.Experimental results show that the recognition accuracy of this method on CK+ and FER2013 data sets reaches 97.88% and 72.81%,respectively,and it has strong competitiveness.In addition,ablation experiments show that the improved method can improve the expression recognition rate,which further verifies its effectiveness.