Light field super-resolution using complementary-view feature attention

Abstract Light field (LF) cameras record multiple perspectives by a sparse sampling of real scenes, and these provide complementary information. This information is beneficial to LF super-resolution (LFSR). Compared with traditional single-image super-resolution, can exploit parallax structure perspective correlation among different views. Furthermore, the performance existing methods are limited as they fail deeply explore across In this paper, we propose novel network, called light complementary-view feature attention network (LF-CFANet), improve LFSR dynamically learning in Specifically, design residual spatial channel module (RCSCAM) effectively interact between Moreover, RCSCAM captures relationships channels, it able generate informative features for reconstructing images while ignoring redundant Then, maximum-difference supplementary branch (MDISB) used supplement from angular positions based on geometric images. also guide process reconstruction. Experimental results both synthetic real-world datasets demonstrate superiority our method. The proposed LF-CFANet has more advanced reconstruction that displays faithful details higher SR accuracy than state-of-the-art methods.