Efficient DenseNet Model with Fusion of Channel and Spatial Attention for Facial Expression Recognition

Facial Expression Recognition (FER) is a fundamental component of human communication with numerous potential applications. Convolutional neural networks, particularly those employing advanced architectures like Densely connected Networks (DenseNets), have demonstrated remarkable success in FER. Additionally, attention mechanisms have been harnessed to enhance feature extraction by focusing on critical image regions. This can induce more efficient models for image classification. This study introduces an efficient DenseNet model that utilizes a fusion of channel and spatial attention for FER, which capitalizes on the respective strengths to enhance feature extraction while also reducing model complexity in terms of parameters. The model is evaluated across five popular datasets: JAFFE, CK+, OuluCASIA, KDEF, and RAF-DB. The results indicate an accuracy of at least 99.94% for four lab-controlled datasets, which surpasses the accuracy of all other compared methods. Furthermore, the model demonstrates an accuracy of 83.18% with training from scratch on the real-world RAF-DB dataset.