Light Field Spatial Super-Resolution via View Interaction and the Fusing of Hierarchical Features

Light field (LF) cameras can capture the intensity and angle information of any scene in one single shot, and are widely used in virtual reality, refocusing, de-occlusion, depth estimation, etc. However, the fundamental limitation between angular and spatial resolution leads to low spatial resolution of LF images, which limits their development prospects in various fields. In this paper, we propose a new super-resolution network based on view interaction and hierarchical feature fusion to effectively improve LF image spatial resolution and preserve the consistency of the reconstructed LF structure. Initially, we provide novel view interaction blocks to represent the relationship between all views, and efficiently combine hierarchical features using a feature fusion block consisting of residual dense blocks (RDBs) to more effectively preserve the parallax structure. In addition, in the process of extracting features, we introduce residual channel-reconstruction blocks (RCBs) to minimize the duplication of channels in the features. An inter-view unit (InterU) and an intra-view unit (IntraU) are employed to minimize redundancy in the spatial domain as well. Our suggested method is tested using public LF datasets, which include large-disparity LF images. The experimental results demonstrate that our method exhibits the best performance in terms of both quantitative and qualitative results.