Abstract
Graph Convolutional Networks (GCNs) have shown strong potential in skeleton-based emotion recognition. However, existing methods often struggle to capture fine-grained details of key body parts. A Hierarchical Channel-Aware Optimization Graph Convolutional Network (HCO-GCN) is proposed. This framework enhances emotion recognition by integrating hierarchical graph decomposition to capture multi-scale dependencies, and the channel-aware optimization mechanism dynamically learns emotion-discriminative spatial patterns through channel wise computations, focusing on fine-grained motion details of critical body joints. To further enhance the representations learned by HCO-GCN, we introduce a multi-stage, multi-level contrastive learning (MMCL) framework. This design encourages more discriminative feature and graph representations across network stages, promoting improved intra-class consistency and inter-class separability. Furthermore, we propose a multi-task learning strategy within a primary–auxiliary task framework, where prior affective knowledge is integrated into the network through affective feature prediction, emotion classification, and contrastive learning, thereby boosting generalization. Experimental results on two public and one private datasets demonstrate the effectiveness of the proposed method in enhancing emotion recognition performance.
Funding Information
  • National Natural Science Foundation of China (62373073)
  • Fundamental Research Funds for the Central Universities (DUT25YG111)

This publication has 44 references indexed in Scilit: