Depth-Guided Full-Focus Super-Resolution Network for Light Field Images

Light field (LF) imaging system captures the two-dimensional (2D) spatial and 2D angular information of scenes within a single exposure time. Due to this distinctive feature, the technique has been rapidly developed over the past two decades. However, the LF images suffer from a low spatial resolution. Currently, numerous deep learning (DL)-based approaches have been employed to address this issue. However, existing super-resolution (SR) networks ignore the defocus blur caused by depth variations, and fail to yield high-resolution (HR) full-focus images by directly processing LF images with depth information. In this paper, to tackle this challenge, we propose a new SR method to reconstruct HR full-focus LF images from low-resolution (LR) multi-defocus LF images. To accomplish this task, The degraded multi-defocus LF dataset is generated by utilizing the depth information intrinsic to LF images as guidance and designing a spatially-variable (SV) degradation method. The method is designed by two parts: a depth-guided image partitioning process and a degradation-prior-SR network. Experimental results have indicated that our method outperforms existing other networks both quantitatively and qualitatively.