Low-Rank Plus Sparse Tensor Models for Light-field Reconstruction from Focal Stack Data

Hand-held light-field cameras have enabled new photographic features such as refocusing and perspective shifts in post-processing. These cameras have traditionally sampled the 4D light-field directly by multiplexing angular measurements with spatial measurements on a single photosensor. This requires an undesirable trade-off between spatial and angular resolution. Focal stack cameras record varying projections of the light-field by capturing a set of photographs at different sensor positions. Prior techniques for reconstructing the 4D light-field from focal stack measurements have required depth estimation or ignored the existence of occlusions in the scene. We present low-rank plus sparse models for the light-field and apply them to the problem of reconstructing from focal stack measurements. We explore regularizers based on low-rank tensor decompositions to better exploit the dimensionality of the data. We optimize our model with a block proximal gradient method using a majorizer that provides a convergence guarantee. Numerical experiments show a several dB improvement in PSNR over traditional reconstruction methods and improved accuracy of depth estimation from light-fields reconstructed by the proposed methods.