ZenNet-SA: An Efficient Lightweight Neural Network with Shuffle Attention for Facial Expression Recognition

Facial expression recognition (FER) is an important task in computer vision, having practical applications in areas such as human-computer interaction, education, healthcare, and online monitoring. Although significant progress has been made possible by deep learning, real-world deployment of the FER system still needs to be improved, particularly with limited resources such as embedded or mobile systems. From this challenging FER task, we propose a lightweight network that combines a lightweight high performance backbone by applying ZenNAS [1] and a distillation feature part after backbone to refine features. Extensive experimental results demonstrate that the proposed ZenNet-SA achieves state-of-the-art results on three FER benchmark datasets, RAF-DB (88.53%) as well as AffectNet 7 class (65.23%) and 8 class (61.32%) in the light-weight models group with a model size less than 1M and computation costs approximately 450M Flops. Our results demonstrate a notable improvement in efficiency, with a decrease of approximately 97% in FLOPS and 91% in the number of parameters compared to POSTER V2 in terms of computational and resource costs.