Advancing Facial Expression Recognition -- Enhanced MobileNetV3 with Integrated Coordinate Attention and Dynamic Kernel Adaptation

This paper presents an improved approach for facial expression recognition (FER), which incorporates the Coordinate Attention (CAM) mechanism into MobileNetV3, a lightweight CNN widely used for its real-time applications on low-power devices. The CA mechanism greatly improves the ability of the model to focus on face regions of interest, as it incorporates positional information, making feature extraction more accurate. Additionally, dynamic kernel adaptation (DKA) and SoftSwish are incorporated into the model to enhance the flexibility and computational efficiency of MobileNetV3. The proposed model was tested in three sets of JAFFE, CK+, and FER2013, where accuracy improvements were reported of 98.84% in the JAFFE dataset, 99.56% on the CK+ dataset, and 88.50% on the FER2013 dataset. These results support the viability and utility of the proposed approach to improve FER, especially in applications that favor higher numerical performance.