Light Field Angular Super-Resolution via Dense Correspondence Field Reconstruction

Abstract

Light field (LF) angular super-resolution (SR) aims at reconstructing a densely sampled LF from a sparsely sampled one. To achieve accurate angular SR, it is important but challenging to incorporate the complementary information among input views, especially when dealing with large disparities. In this paper, we propose to reconstruct dense correspondence field among different views for LF angular SR. According to the LF geometry structure, we first capture correspondences along the horizontal and vertical axes of input views with a global receptive field. We then incorporate the linear structure prior among angular viewpoints to reconstruct a dense correspondence field. With the reconstructed dense correspondence field, the relationship between each target view and the input views is constructed. Next, we develop a view projection approach to project input views to the target positions. Moreover, a projection loss is introduced to preserve the LF parallax structure. Extensive experiments demonstrate that our proposed network can recover accurate details and preserve LF parallax structure. Comparative results show the advantage of our method over state-of-the-art methods on synthetic and real-world datasets.