基于注意力多尺度融合的人脸表情识别算法研究

Research on Facial Expression Recognition Algorithm Based on Attention Multi-Scale Fusion

信息技术在教学中的应用导致师生之间缺乏一定程度的情感交流,为了弥补授课过程中的情感缺失,获得更好的教学反馈,提出基于注意力机制与多尺度特征融合(ASMF)的人脸表情识别算法。该算法以Resnet 50作为骨干网络,首先通过对多层卷积神经网络的输出特性进行多尺度的融合,引入上下文信息的同时提取更加丰富有效的表情特征信息;其次将注意力机制融入网络中,通过对各通道进行加权学习,得到注意力特征图,从而增强特征的表达能力,抑制冗余信息的影响;然后加入Dropout机制和Softmax Loss损失函数,进一步提高提取到的表情特征的可判别性;最后,利用消融试验在公开的数据集与自制的学生课堂表情数据集上验证该算法的有效性和稳定性,识别准确率达到93.87%。

The application of information technology in teaching leads to a lack of emotional communication between teachers and students.In order to compensate for the lack of emotional communication during the teaching process and obtain better teaching feedback,a facial expression recognition algorithm based on attention mechanism and multi-scale feature fusion(ASMF)is proposed.The algorithm uses Resnet 50 as the backbone network.It firstly fuses the output characteristics of multi-layer convolutional neural networks at multiple scales,introduces contextual information while extracting richer and more effective expression feature information.Secondly,the attention mechanism is integrated into the network,and through weighted learning of each channel,attention feature maps are obtained to enhance the expression ability of features and suppress the impact of redundant information.Then,the Dropout mechanism and Softmax Loss function are added to further improve the discriminability of the extracted facial features Finally,the effectiveness and stability of the algorithm are validated by using ablation experiments on both publicly available datasets and self-made student classroom expression datasets,with a recognition accuracy of 93.87%.