To address the problems of occlusion and postural change in pedestrian reidentification and the low identification rate of current networks, a multibranch pedestrian reidentification network model with different spatial dimensions is proposed. First, IBN-Net50-a serves as the backbone network to extract features. Second, the last two convolution layers are fused with the batch feature-discarding method to enhance the local focal feature learning. Finally, features with different dimensions can be combined to obtain additional valuable information at both the shallow and deep levels. The combined strategy of triplet loss and label smoothing losse is adopted in network training. Further, three commonly used benchmark datasets, Market-1501, DukeMTMC-reID, and CUHK03, are used for experimental verification; the datasets are divided based on the mainstream strategy. Experimental results show that the proposed method achieves an effective feature generalization capability. On the Market1501, DukeMTMC-reID, and CUHK03 datasets, Rank-1 and mean average precision (mAP) values of 95.3% and 86.8%, 88.5% and 75.9%, and 80.9% and 77.8%, respectively, are achieved.