Unconstrained palmprint recognition has shown great potential for practical applications due to its safety, convenience, and hygiene. However, unconstrained palmprint recognition often has difficulties in existing region of interest (ROI) and robust feature descriptor. In this paper, we propose a vision transformer network MDHs-ViT for unconstrained palmprint recognition based on defining the enlarged ROI (EROI) of palmprint images. Firstly, EROI breaks the limitation of traditional small ROI extraction of palmprint, that is, without aligning, EROI can still bring better performance for unconstrained palmprint recognition. Then, according to the inherent traits of EROI, the proposed MDHs-ViT fully learns discriminative features by multi-scale feature learning and dense attention heads transformer. Extensive experimental results on three public datasets demonstrate the effectiveness of the defined EROI and the proposed MDHs-ViT.