Abstract: In order to increase the robustness of finger vein descriptors and reduce the number of network parameters, we proposed a way of modifying the VGGFace-Net and using the Vector of Locally Aggregated Descriptors(VLAD). The number of parameters of our obtained network is only 0.3M. The VLAD can cluster and rearrange the local descriptors, which makes our descriptors more robust against the changes of finger posture. Since the size of public finger vein databases is small, we trained the network by the triplet with the hard negative sample mining strategy. However, the Triplet Loss does not constrain the intra-class variance of the sample pair distance, we further proposed an improved loss called Pair-center-constrained loss. By treating positive and negative sample pairs as two categories, we can drive them even further to their class centers, which can increase the intra-class compactness. Experimental results on three public databases FV-USM, SDUMLA, and MMCBNU show that the proposed method is better than two state-of-the art methods in terms of accuracy. Meanwhile, our finger vein descriptors have better robustness against random translation.