Application of lightweight facial expression recognition based on deep convolutional asymmetric residual attention network model in education management

In view of the low accuracy of facial expression recognition in natural state, which is easily affected by noise and other factors, this paper proposes a lightweight facial expression recognition method based on deep convolutional asymmetric residual attention network model. In this paper, Ghost module and depth-separable convolution are introduced to replace 1×1 convolution and 3×3 convolution in Bottleneck, which retains more original feature information and improve the feature extraction ability of backbone branches. A parallel deep convolutional residual structure is designed in the shallow network to enhance the model’s ability to represent the local details of facial expressions and integrate them with the global features. A spatial grouping enhanced attention mechanism is established in the deep network to improve the stability of expression feature distribution and enhance the ability of the model to discriminate subtle changes in expression. In order to avoid over-fitting of the model, the output structure of the backbone network is improved without greatly increasing the computational complexity. The Mish activation function is used to replace the ReLU activation function in Bottleneck, which improves the accuracy of expression recognition. The expression recognition accuracy of the proposed method on the published seven-classification dataset RAF-DB, AffectNet-7 and eight-classification dataset AffectNet-8 reaches 88.44%, 63.20% and 60.23%, respectively. The experimental results show that the proposed method can improve the expression recognition accuracy while reducing network parameters. It is proved that the method is effective and has a certain application prospect.