Attention and Relative Distance Alignment for Low-Resolution Facial Expression Recognition

In real-world scenarios, facial images obtained by many devices often exhibit low resolution. However, the performance significantly degrades when we apply the existing methods in low-resolution facial expression recognition. Therefore, addressing the problem of low-resolution images in facial expression recognition becomes an important undertaking. Previous attempts to tackle this problem have been limited. For this, we propose a novel Attention and Relative Distance Alignment (ARDA) method by integrating knowledge distillation in low-resolution facial expression recognition. Specifically, the Attention Alignment module guides the student model to focus on the most crucial region of the facial image by enabling the low-resolution student model to learn the attention map of the high-resolution teacher model. The Relative Distance Alignment module utilizes the relative distance between facial image features to transfer differences between different low-resolution facial images from the teacher model to the student model, helping the student model better grasp the differences between expressions. Extensive experiments have shown that the ARDA method effectively transfers knowledge from high-resolution teacher model to low-resolution student model, achieving state-of-the-art performance in synthetic low-resolution facial expression recognition datasets.