An improved finger vein recognition method based on ResNet is proposed to solve the problem of finger vein extraction difficulties and insufficient recognition accuracy. First of all, depthwise over-parameterized convolution (DO-Conv) is used to replace the traditional convolution in the network, while reducing the model parameters and improving the network recognition rate. Then, the spatial attention module (SAM) and squeeze-and-exception block (SE-block) are fused and applied to an improved ResNet to extract the detailed features of the image in the channel and spatial domain. Finally, label smoothed cross entropy (LSCE) loss function is used to train the model in order to automatically calibrate the network to prevent errors in classification. The experimental results show that the improved network model is not easily affected by the image quality. And the recognition accuracy of FV-USM and SDUMLA can reach 99.4919% and 99.4485%, respectively, which is significantly higher than that of the previous network. Compared with other models, the proposed method has a significant improvement on the recognition accuracy.