Face expression recognition via cross-scale cross-attention and dynamic token pruning
Yongwei Nie, Rong Pan, Xu Cao · 2025
Facial expression recognition (FER) is a key task in computer vision, with broad applications in education, healthcare, and human-computer interaction. Despite significant progress driven by deep learning, FER still faces challenges such as the loss of low-level details during the hierarchical abstraction process in deep networks, making it difficult to detect subtle facial expression changes. While multi-scale architecture helps retain local details, it often lack effective feature interaction and suffer from feature redundancy. To address these issues, this paper proposes a FER method based on cross-scale cross-attention and dynamic token pruning. The method extracts local details and global semantics through a multi-scale feature architecture, and a cross-scale cross-attention mechanism is designed to guides high-level features to actively retrieve low-level details, enhancing feature collaboration. Additionally, the dynamic token pruning strategy removes low-importance tokens to balance accuracy and efficiency. Experimental results demonstrate that the proposed method outperforms existing approaches on multiple FER datasets, validating its effectiveness.