Explicit Spatial-Angular Pattern Embedding for Heterogeneous Imaging-Oriented Light Field Spatial Super-Resolution

Light field (LF) imaging often faces a severe trade-off between spatial and angular resolution due to limited sensor capacity. Although various LF spatial super-resolution (SR) approaches have emerged, they often struggle with blurred textures and degraded parallax structures at large upscaling factors (e.g., 8×). Heterogeneous imaging aims to addresses this by leveraging supplementary 2D high-resolution (HR) information; however, effectively bridging 2D and LF data for angular-coherent high-frequency transfer remains challenging. In this paper, we propose a novel learning-based framework using explicit spatial-angular pattern embedding for heterogeneous LF spatial SR. Specifically, the proposed method first develops a pseudo-LF generation and implicit alignment module to establish the correspondence between the 2D HR image and the low-resolution (LR) LF image. Subsequently, local spatial and global angular pattern estimation strategies are designed to exploit the intrinsic LF geometry. Based on this, a spatial-angular pattern-embedded LF feature enhancement module is constructed to generate detailed LF texture features, thereby enabling high-fidelity SR LF image reconstruction. To handle excessive resolution gaps, a progressive restoration scheme is introduced, incorporating weight-sharing across different upscaling scales to fully exploit the network’s learning potential. Experimental results on multiple benchmark datasets show that the proposed method generates super-resolved LF images with sharp spatial textures and accurate angular parallax, outperforming state-of-the-art methods in both objective evaluation and subjective quality. Our code is available at https://github.com/YeyaoChen/LFSSR-ESAPE.