Vol. 56, Issue 1, pp. 149-164 (2026)

Abstract

Based on the two-plane representation model of the light field, light field cameras can synchronously record positional and directional information of the rays by sacrificing spatial resolution. To improve the spatial resolution of light field images, this paper presents a light field super-resolution network built upon the mechanism of multi-dimensional disparities mining. In the shallow layers of the network, a multi-scale self-attention module is designed, which works in conjunction with a residual atrous spatial pyramid pooling module to extract features from sub-aperture images. For the deep feature extraction, a multi-dimensional feature separator-mixer is developed to mine disparities from different forms of light field subspace, and an iterative structure is employed to extract deep fused disparity among different sub-aperture images. At the end of the network, a global-local feature refine module is added to further improve the quality of the output images. By adopting the mechanism of disparity iterative separation-fusion and re-enhancement, the proposed network achieves high-quality super-resolution reconstruction of sub-aperture images with a relatively compact architecture. Experimental results verify the superiority of the proposed network. In particular, with only 3.83 M parameters, the network achieves 4 times super-resolved images with an average PSNR of 32.27 dB in 5 public datasets.