MD-GLFNet: multi-directional global–local feature fusion network for facial expression recognition

Facial expression recognition is a fundamental task in computer vision. As facial expressions arise from interactions among multiple facial regions, accurate recognition relies on both local details and cross-region relationships. To address these challenges, we propose a Multi-Directional Global– Local Feature Fusion Network (MD-GLFNet). First, we propose a Multi-Directional Gabor Feature Enhancement (MDGFE) module that extracts and fuses multi-directional local features, thereby enhancing detailed texture representation. Second, a Global– Local Spatial Interaction (GLSI) module is presented that enables multi-region and multi-scale spatial interactions by introducing global tokens. Furthermore, we design a Bidirectional Channel– Spatial Modulation (BCSM) module with two residual branches—one that applies channel modulation followed by spatial modulation, and another that applies spatial modulation followed by channel modulation—for joint channel– spatial feature modeling. Finally, a lightweight Token Pooling Classification (TPC) head is designed for improved discriminative ability and reduced computational cost. The proposed method achieves recognition accuracies of 90.29%, 74.37%, 88.76%, and 96.67% on RAF-DB, FER2013, FERPlus, and JAFFE, respectively. Comparative results demonstrate that our method achieves competitive performance across different datasets. Visualization analyses further show that the proposed method focuses more accurately on expression-related facial regions. Moreover, all these benefits are achieved while keeping computational overhead low. Ablation studies further confirm that each module provides consistent performance gains across different datasets, verifying the effectiveness of the proposed method.