A TV regularisation sparse light field reconstruction model based on guided-filtering

Joint Fourier Disparity Layers Unrolling With Learned View Synthesis for Light Field Reconstruction From Few-Shots Focal Stacks

“…Classical approaches to solve the optimization problem posed in Eq. ( 2) use iterative optimization algorithms, usually introducing hand-crafted [11] or pre-learned [14], [15] image priors. Unrolled optimization methods have enabled major progress in the field of image inverse problems.…”

“…Handcrafted priors have then been introduced in the formulation of the inverse problem of recovering light fields from focal stacks. Gao et al [11] proposed the ADMM algorithm with a TV-regularization along with a guided filter. A convolution kernel is derived to model the focal stack imaging process.…”

“…several images of the scene at different focus distances, in order to reconstruct a light field. However, existing reconstruction methods [10], [11] typically require focal stacks with dense sampling in the focus dimension, so that the details can be retrieved at every depth in the scene. Hence, many shots are needed in the capture process.…”

“…While traditional approaches consider hand-crafted priors to regularize the optimization problem, such as sparsity [12], smoothness [13] or total variation [11], significant advances have been achieved thanks to the introduction of learned priors. A first category of methods, referred to as "Plug-and-play" (PnP) [14], [15], has been introduced where a pre-learned prior is plugged into an iterative optimization algorithm.…”

“…In the context of light field reconstruction from focal stack measurements, optimization algorithms that have been designed in the literature mostly use handcrafted priors [11], [25]- [29]. One of them uses a Tikhonov regularization to optimize the Fourier Disparity Layers (FDL) representation of a light field [29], which has been introduced as a compact representation of scenes.…”