Abstract: Multi-scale features are critical in dense prediction tasks within the point cloud domain. Existing 3D point cloud processing techniques predominantly rely on encoder-decoder frameworks, which extract and integrate multiscale features via a backbone network. However, these methods often employ delayed fusion strategies, resulting in insufficient feature integration. To address this issue, this paper introduces a novel high-resolution feature network for 3D point cloud, named HRFN3D, specifically for point cloud classification and segmentation tasks. HRFN3D innovatively employs a relational learning module to perform feature fusion at an early stage, facilitating interactions between low-resolution high-semantic points and high-resolution low-semantic points. This early fusion ensures that high-resolution points retain semantic information from the outset, facilitating subsequent feature learning. In the later stage, the global feature vectors are generated by combining different pooling strategies and spliced with the original point features, preserving the details and enhancing the representation of the global features. The experimental results show that HRFN3D improves the Class mean and Instance mean Intersection over Union by 2.2 percentage point and 0.9 percentage point, respectively, and achieves the average class ratio of 86.3%. On the ModelNet40 data set, our proposed method achieves the highest class average accuracy of 91.5% with 4.3M parameters. These results validate the effectiveness of HRFN3D in multi-scale feature processing.