Hierarchical spatial–angular integration for lightweight light field image super-resolution

Recent advancements in light field super-resolution (LFSR) using Transformer-based methods have shown promising improvements. However, existing methods often restrict the modeling of spatial and angular information to individual sub-aperture images (SAIs) and macro-pixel images (MacPIs), limiting the potential for comprehensive information extraction. This results in a common trade-off between model complexity and performance. Additionally, the absence of lightweight models hinders the practical application of LFSR. To address these challenges, we propose a lightweight Transformer-based network that efficiently integrates spatial and angular representations while extending the modeling scope across both SAIs and MacPIs at a low complexity cost. Our approach introduces two key innovations: hierarchical spatial integration (HSI) and hierarchical angular integration (HAI), which capture intra- and inter-spatial and angular dependencies within SAIs and MacPIs, respectively. We then design a third-order loop structure that combines these two integrations, facilitating the progressive merging of spatial and angular information without introducing extra parameters. To further enhance model performance, we incorporate a dedicated epipolar plane image (EPI) module to capture the geometry of the spatial–angular correlation. Extensive experimental results demonstrate that our proposed model significantly outperforms state-of-the-art LFSR algorithms, achieving superior visual quality while maintaining lower computational complexity.