LGENet: A Novel Feature Extraction Network for Point Cloud Registration in 3D Reconstruction

Point cloud registration is mainly achieved through traditional methods and learning-based methods. For learning-based methods, the feature extraction has a great influence on the registration performance. This paper proposes a local-global enhancement network (LGENet) for learning-based methods. LGENet consists of adaptive point graph convolution (APGConv) and convolutional block attention module (CBAM). The APGConv effectively fuses the local features extracted by adaptive graph convolution and the global features extracted by point convolution. The CBAM further enhances the learned point cloud features. In the comparison with various point cloud registration models, the LGENet model demonstrates superior performance, as validated on the ModelNet40 dataset. The LGENet model demonstrates a significant advantage in rotational error (RMSE(R) and MAE(R)), while also achieving satisfactory performance in translation error (RMSE(t) and MAE(t)). In addition, the ablation experiments validate the effectiveness of both the APGConv and CBAM modules.