Mamba-Based Global Correlation Learning for Light Field Spatial Super-Resolution

Light field spatial super-resolution is a challenging task due to complex 4D light field structure. How to explore the global correlation of the light field to improve the reconstruction quality remains a key problem in LF-SSR. In this paper, a novel light field spatial super-resolution method with Mamba-based global correlation learning is proposed, which effectively explores pixel-wise correlation within and across different subspaces based on Mamba block. Specifically, a tri-subspace scanned feature enhancement module is proposed to capture the long-range dependency within each 2D light field subspace with the specially-designed scanning schemes. Besides, a Mamba-based spatial-angular information fusion module is designed to mine the global complementary information among the spatial, angular, and EPI features for light field image reconstruction. Extensive experiments show that the proposed method is superior to other advanced light field super-resolution methods.