Abstract: Aimed at the problem of low 3D face recognition performance under the condition of lack of 3D face resources, a 3D face recognition method fused with data augmentation structure and end-to-end network is proposed. The method consists of two parts, the end-to-end learning network IMPNet and the 3D face data augmenting structure. It used Pointnet++ as the backbone network, and a multi-scale feature average add-drop fusion strategy and SA network module reconstruction was proposed, which increased the inference speed by 37.7%. By improving the SSM, the data augmentation structure based on the optimized deformation statistical model GSSM was proposed for the first time, so that the network with better performance could be obtained without a large number of real data as the training set. The accuracy of this method on public datasets FRGCv2 and Bosphorus is 98.9% and 99.1%, which is a great improvement compared with the current mainstream methods.