Improved Facial Emotion Detection Using Coordinate Attention

Facial expression recognition plays a crucial role in many real-world applications. Current methods primarily rely on Vision Transformers and CNNs, but both have limitations: Vision Transformers demand extensive computational resources, while CNNs struggle to capture long-range dependencies. To overcome these challenges, we propose a new facial emotion recognition model based on the FERVT framework. Our model replaces the high-level feature extraction component of FERVT with Coordinate Attention, resulting in a more lightweight architecture with improved spatial information capture. We also refine the feature map fusion process to extract more detailed information and introduce a new loss function to enhance model training. Our model achieves an 8% improvement in accuracy across multiple public datasets compared to the FERVT method. It also outperforms VGGI6, ResNetl8, and GoogleNet by an average of 12% without pretraining, and by 5% on average when compared to pretrained models.