The acquired point clouds are generally incomplete because of the existing problems of low resolution and view occlusion in 3D scans. Focusing on these issues, this paper proposes a multi-scale point cloud shape completion network, which predicts the missing cloud points and recovers the complete geometric shapes. The network achieves multiple scales of point clouds by subsampling incomplete point clouds and predicts point clouds at multiple scales through a feature pyramid decoder. Then it calculates the loss of predicted point clouds at each level to ensure the generation of more realistic point clouds. Our major contributions lie in two aspects: on the one hand, we design a feature extraction module called composite multi-layer perceptron to improve the fusion of local and global information comprehensively. On the other hand, we introduce a patch discriminator into the network, which better controls the details of the generated point cloud by dividing the generated into multiple local regions and discriminating independently. Experimental results on the ShapeNet subset show that compared with the mainstream PF-Net structure, the accuracy of the network generated point cloud is improved by 2.9%, and the coverage of the generated point cloud is increased by 6.3%.