Improved ConvNeXt Facial Expression Recognition Embedded with Attention Mechanism

Facial expression recognition (FER) is an emerging and important research field in the field of pattern recognition, with wide applications in safe driving, intelligent monitoring, and human-computer interaction. This article addresses the problems of insufficient key information extraction, low recognition accuracy, and easy overfitting in facial expression recognition, and proposes an ECA-ConvNeXt network based on transfer learning strategy and channel attention mechanism. Firstly, the weights of the pre-trained model are initialized using transfer learning on the FER 2013 dataset. Secondly, a series of data augmentation operations are performed on the facial images, allowing them to pass through the ECA-Net attention module of the network, enhancing the key information of the feature regions with high relevance to expressions and suppressing the interference of irrelevant regions in the feature maps. Finally, the inverse bottleneck layer, maximum pooling layer, global average pooling layer, and classification layer are sequentially passed into the network to accelerate the convergence speed and improve the expression recognition rate. Compared to the baseline network, the improved network achieved an accuracy of 72.86%, a recall rate of 72.04%, and a specificity of 64.15% on the FER 2013 dataset. Compared to the commonly used ResNet network and its improvement methods, the proposed ECA-ConvNeXt in this article achieved a 0.19% improvement in recognition accuracy.