<p>Most existing point cloud upsampling methods primarily rely on feature expansion to increase point cloud density, which limit their performance in generating high-resolution point clouds. To address this issue, we propose a self-supervised progressive network for point cloud upsampling (SPU-PRTI). The network, alongside a hierarchical supervision framework, achieves high-fidelity point cloud surface reconstruction by designing generation module and refinement module to achieve high-fidelity surface reconstruction.Specifically, our method first employs a two-stage interpolation strategy within the generator to capture the topological structure of the input point cloud, generating an initial dense point cloud. Subsequently, the two-stage interpolated point clouds are associated with folded feature expansion to refine surface uniformity and structural details. Furthermore, a Point Gated Recurrent Unit (PGRU) is introduced in the refinement stage to utilize the memory characteristics of relative point positions for guiding point movements. Finally, these modules are integrated into a hierarchical supervision architecture, where collaborative optimization across hierarchical supervision ensures high-fidelity point cloud reconstruction. To validate the effectiveness of our approach, we conducted qualitative and quantitative experiments on both synthetic datasets and real-world scanned datasets. Experimental results on synthetic data demonstrate that SPU-PRTI outperforms baseline methods, achieving lower Chamfer distance (0.54), Hausdorff distance (5.35) and point-to-surface distance (2.70) while preserving local details and reducing outliers.</p>