Abstract: From the existing normal estimation studies, how to select a proper neighbor of the target point is a key and difficult problem, especially when there exists high level of noises in the target point cloud. In order to improve the accuracy of the normal estimation, a new network architecture is proposed to estimate the normal of the point clouds, which can adaptively selects the neighbors, keeps the features and resists the noises. First, a new bilateral non-local feature enhancement module is introduced to improve the network’s ability to learn the local geometric features of the model, which can adaptively selects the neighbors according to the pre-learning features of the network and the geometric characteristics of the neighbors and then enhances the local features. Then, the combination of local features and global features is used to describe the complete geometric features of the point cloud, based on which further local surface fitting and normal estimation are carried out. Finally, in the local surface fitting, a feature-preserving neighbor loss is proposed, which can adjust the fitting weight of the neighboring points according to the degree of noise interference of the neighboring points, to realize the local surface fitting with the feature of detail preservation and improve the robustness of the method to noises. The PCPNET dataset is used to train and test the models. From extensive experiments, compared with the most related state-of-the-arts, proposed method gets better normal estimations results for the different noise levels and density distributions of the point clouds compared with existing methods, and it can also promote surface reconstruction and other point cloud processing tasks.