Light Field Super-Resolution Based on Spatial and Angular Attention

Light field (LF) images captured by LF cameras can store the intensity and direction information of light rays in the scene, which have advantages in many computer vision tasks, such as 3D reconstruction, target tracking and so on. But there is a trade-off between the spatial and angular resolution of LF images due to the fixed resolution of sensor in LF cameras. So LF image super-resolution (SR) is widely explored. Most of the existing methods do not consider the different degree of importance of spatial and angular information provided by other views in LF. So we propose a LF spatial-angular attention module (LFSAA) to adjust the weights of spatial and angular information in spatial and angular domain respectively. Based on this module, a LF image SR network is designed to super-resolve all views in LF simultaneously. And we further combine the LF image SR network with single image SR network to improve the ability to explore spatial information of a single image in LF. Experiments on both synthetic and real-world LF datasets have demonstrated the performance of our method.