Facial expressions are a complex form of biological motion involving dynamic configurations of facial muscle movements that encode affective states, cognitive processes, and social intentions. Micro-expressions as a subset of facial expressions are particularly valuable for emotion analysis due to their involuntary nature, brief duration, and high truthfulness. This paper proposes an Enhanced Multi-scale Hierarchical Network (EMHNet) for micro-expression recognition. Our architecture introduces a Hierarchical Mixture of Experts (HMoE) system employing specialized transformers for four critical facial regions coupled with a global transformer for holistic integration and an adaptive multi-scale framework featuring dynamic block partitioning and cross-scale attention gates for optimized feature extraction. Experimental evaluation is performed on SMIC, CASME II and SAMM benchmarks and a composite dataset of the three datasets. Experimental results show that our proposed EMHNet achieves competitive performance compared to existing state-of-the-art methods, demonstrating its effectiveness.