Detail-Preserving Transformer for Light Field Image Super-resolution

Recently, numerous algorithms have been developed to tackle the problem of light field super-resolution (LFSR), i.e., super-resolving low-resolution fields gain high-resolution views. Despite delivering encouraging results, these approaches are all convolution-based, and naturally weak in global relation modeling sub-aperture images necessarily characterize inherent structure fields. In this paper, we put forth a novel formulation built upon Transformers, by treating LFSR as sequence-to-sequence reconstruction task. particular, our model regards each vertical or horizontal angular view sequence, establishes long-range geometric dependencies within sequence via spatial-angular locally-enhanced self-attention layer, which maintains locality image well. Additionally, better recover details, propose detail-preserving Transformer (termed DPT), leveraging gradient maps guide learning. DPT consists two branches, with associated for learning from an original sequence. The branches finally fused obtain comprehensive feature representations reconstruction. Evaluations conducted on number datasets, including real-world scenes synthetic data. proposed method achieves superior performance comparing other state-of-the-art schemes. Our code is publicly available at: https://github.com/BITszwang/DPT.