Cross-Domain Mesh–Progressive Fusion for Light Field Reconstruction

Abstract

Light field angular super-resolution (LFASR) seeks to recover missing viewpoints from sparsely captured angular samples, while preserving the underlying scene structure across views. This task is complicated by disparity-induced variations, incomplete visibility near occlusion boundaries, and view-dependent appearance changes caused by reflective surfaces. In this work, we propose a cross-domain EPI–spatial–angular (X-ESA) framework that combines EPI-based geometric modeling with complementary spatial and angular feature extraction. A CSWin-based EPI transformer captures long-range, slope-aligned dependencies, while dedicated convolutional branches refine spatial details and cross-view relationships. These features are integrated through a mesh–progressive fusion strategy, followed by channel attention and spatial residual refinement. Experiments on real-world and synthetic datasets under the 2×2→7×7 setting result in average performances of 41.29 dB/0.989 SSIM and 40.98 dB/0.989 SSIM, respectively. The proposed model contains 2.84M trainable parameters and provides competitive reconstruction quality relative to existing LFASR methods.