Face Expression Recognition Based on Lightweight Fused Attention Mechanism

Abstract

To address the problems of redundant model counts, large computational effort, poor targeting of effective feature extraction and easy loss of large amount of information in traditional convolutional neural networks for expression recognition, a lightweight fused attention mechanism approach for face expression recognition is proposed. The method is based on ResNet convolutional neural network, and the depthwise separable convolutional module is added in the feature extraction stage to reduce the number of parameters, and then the attention mechanism of the fusion channel is used to improve the extraction and representation ability of the model for important feature information. The PReLU is used to replace the ReLU to prevent Dying ReLU problems. The model has been simulated on the public RAF-DB dataset. The results show that the accuracy of facial expression recognition reached 85.53%, while the number of parameters and computational effort are kept at low levels. The results verify the effectiveness and superiority of the improved model.