Infant and toddler facial expression recognition based on topology-aware pooling graph U-Net

Abstract

Facial expressions are essential for infants and toddlers to communicate emotions and interact socially. Accurate recognition of these expressions is vital for understanding their needs and enhancing adult-child communication. However, this task is challenging due to the subtlety of infant expressions, environmental interference, and the inherent small inter-class but large intra-class differences in early emotional development. To address these challenges, this paper proposes a Topology-Aware Pooling Graph U-Net for infant facial expression recognition in natural scenes. This is the first work to introduce Graph U-Net into this domain. The model constructs graph structures to capture similarity relationships among feature vectors, enabling deep exploration of topological information in high-dimensional features. A novel topology-aware graph pooling mechanism is designed to effectively assess node importance within the graph. Experimental results on the IFER dataset show that the proposed method achieves 93.22% accuracy, outperforming state-of-the-art approaches. It excels in distinguishing visually similar expressions such as "startle" (90.41%) and "crying" (95.03%), demonstrating its robustness and effectiveness.