DRFusion: Enhancing balanced and sufficient multimodal learning for human emotion recognition | Synapse