Sparse-to-dense light field reconstruction based on Spatial–Angular Multi-Dimensional Interaction and Guided Residual Networks

Dense light fields contain rich spatial and angular information, making them highly valuable for applications such as depth estimation, 3D reconstruction, and multi-view elemental image synthesis. Light-field cameras capture both spatial and angular scene information in a single shot. However, due to high hardware requirements and substantial storage costs, practical acquisitions often yield only sparse light-field maps. To address this problem, this paper proposes an efficient end-to-end sparse-to-dense light-field reconstruction method based on Spatial–Angular Multi-Dimensional Interaction and Guided Residual Networks. The Spatial–Angular Multi-Dimensional Interaction Module (SAMDIM) fully exploits the four-dimensional structural information of light-field image data in both spatial and angular domains. It performs dual-modal interaction across spatial and angular dimensions to generate dense subviews. The channel attention mechanism within the interaction module significantly improves the image quality of these dense subviews. Finally, the Guided Residual Refinement Module (GRRM) further enhances the texture details of the generated dense subviews, enhancing the reconstruction quality of the dense light field. Experimental results demonstrate that our proposed network model achieves clear advantages over state-of-the-art methods in both visual quality and quantitative metrics on real-world datasets. • Spatial-angular multi-dimensional interaction for dense light field reconstruction. • Employs channel attention to enhance information use and sub-view quality. • Designs guided residual refinement to optimize dense light field quality. • Experiments on multiple datasets show superior PSNR/SSIM and visual quality.