Unsupervised person re-identification was attracted much attention due to its good scalability in real surveillance scenarios. The existing unsupervised person re-identification methods mainly trained the network by obtaining rough global features through the basic backbone network, and rarely used local refinement branches and global feature sharing to form more discriminative feature descriptors. This paper proposed a feature extraction network based on local refinement multi-branch and global feature sharing. The network combined rough global features and fine features in local refinement branches to obtain diverse feature expressions of person. In addition, in order to improve the ability of the branch network to capture information of potential key areas, an attention block of channel refinement information fusion was placed before the branch operation to enhance the network’ s attention to person features and perform dedicated learning of refined features. The experimental results on Market-1501, DukeMTMC-reID and MSMT17 datasets verified the effectiveness of the proposed method. The average accuracy (mAP) was improved by 4. 4%, 3. 2% and 6. 4% respectively, and the average accuracy on Market-1501 dataset reached 83. 3%.