An Efficient Multi-view Transformer for Emotion Recognition of University Students

Accurate emotion recognition for university students is essential for mental health monitoring and engagement analysis, yet the intensive computational cost of standard Vision Transformers hinders their deployment on resource-constrained edge devices. To address this challenge, we propose an Efficient Multi-view Transformer (EFFormer) designed for real-time affective computing in campus environments. EFFormer first employs a Bidirectional Mamba strategy to synthesize unified affective representations from multiple views, effectively capturing complex cross-view correlations with linear complexity. Furthermore, we introduce an instancespecific adaptive gating mechanism that dynamically executes patch pruning, attention head activation, and transformer block skipping based on the complexity of each input sample. By jointly optimizing the backbone with a resource-aware loss function and utilizing Gumbel-Softmax reparameterization, EFFormer achieves a superior trade-off between recognition accuracy and inference efficiency. Experimental results demonstrate that our framework significantly reduces computational overhead and latency while maintaining high-fidelity emotional state recognition, providing a practical and robust solution for intelligent emotion monitoring in university settings.