Spatial-angular enhanced network for light-field image super-resolution with geometry-assisted upsampling

Light field (LF) cameras can capture spatial and angular information of 3D scenes, enabling applications in virtual reality and 3D reconstruction. The development of the LF imaging technique has been hampered by the trade-off inherent in the angular and spatial resolution of LF cameras. To solve this challenge, we propose a content-enhanced network. We explore the texture information of complex scenes in LF to recover the texture details more effectively. Our proposed network contains two stages. In the first stage, we propose a local texture exploration network and a spatial-angular information enhancement network. With the local texture exploration network, we extract the gradient maps as a guide to exploring local texture features. In addition, we design a spatial-angular information enhancement network to disentangle and enhance the LF features. In the second stage, we introduce a geometry-assisted upsampling block that enhances LF spatial super-resolution by leveraging geometric information. Our method is capable of recovering more texture details while simultaneously maintaining excellent geometric consistency during the reconstruction of LF images. Comprehensive experiments have confirmed that the proposed method achieves superior performance.