Micro expressions are spontaneous, fast, and subtle, difficult to fabricate or suppress, and therefore are important nonverbal cues for revealing true emotions. Existing deep learning methods have bottleneck problems in multidimensional feature modeling: the singular aggregation method in scale dimensions limits the fusion of multilevel information, fixed-position encoding strategy struggles to adapt to complex facial geometry in the spatial dimension, and treating all channel dimensions equally overlooks the varying importance of emotion-related features. To address these issues, this paper proposes a hierarchical adaptive visual enhancement transformer network with attention and distance correlation (HAVETNet) as a unified framework for multidimensional modeling in micro-expression recognition tasks. HAVETNet is based on the hierarchical Transformer and comprises three key components: the enhanced multi-head self-attention mechanism (E-MSA), the 2D learnable axial positional embedding (2DLAPE), and the multi-scale parallel attention-guided channel aggregation (MACA). Specifically, the MACA module employs multi-scale parallel convolutions and channel attention to integrate emotional information from multiple receptive fields and adaptively enhance features, improving the network’s ability to represent features at different scales. The E-MSA module efficiently extracts emotion-related features while suppressing redundant information via channel attention. The 2DLAPE module introduces learnable row and column position embeddings in the spatial domain to better capture the local semantic structure of facial geometry, further enhancing representation capacity. Experiments on CASME II, SAMM, SMIC, and their composite datasets demonstrate that the proposed approach outperforms most existing methods, generating highly discriminative features and achieving significant performance gains.