With the rapid development of computer 3D vision, point cloud data containing spatial geometric information is widely used in robots, autonomous driving and other scenes. However, due to occlusion, angle limitation and other reasons, geometric information is often missing. In order to solve this problem, SRC-Net is proposed. Firstly, a geometric skeleton is reconstructed from the incomplete point cloud using a designed skeleton reconstruction network that fuses a dynamic graph convolutional network encoder and a folding network decoder, and then the auto-encoder structure is used to establish the mapping from the geometric skeleton to the uniform and complete point cloud. Finally, the completion results on the MVP dataset show that SRC-Net can generate high-quality complete point clouds that are more evenly and smoothly distributed than existing completion networks, and can achieve more detailed completion effects. It provides a new idea and method for point cloud deep learning completion, and has certain guiding significance.