Abstract: 3D face recognition is an important part of the field of computer vision. Pointnet relies on deep learning to solve the disorder of point clouds and realize the global feature extraction. However, due to the lack of detailed texture of point clouds, it is difficult to realize face recognition in complex situations only by global features. In deal with the above problems, a local feature descriptor is proposed to describe the local spatial geometric features of the point clouds, and a key feature enhancement mechanism is introduced to enhance the key features of the face through the probability distribution, which can reduce the interference of unnecessary features and effectively improve the accuracy of the model. Experiments were carried out on public data sets CASIA-3D, Lock3DFace and Bosphorus. The results show that our method can deal well with the change of expression, partial occlusion and interference of head pose, especially in weak light conditions, compared with RP-Net, the accuracy is improved by 1.1 percent, and the method also has good real-time performance.