Light field reconstruction from 2D projections by compressive sensing

Abstract

Light field photography has gained popularity recently due to its imaging capabilities. The light field is a technique that stores the spatial and angular information of an optical signal from a three-dimensional object in a four-dimensional (4D) representation. The acquisition of high-resolution light fields is still a challenge, although there exist commercial light field cameras, those devices trade spatial resolution by offering different views of the same scene. A novel approach to obtain high-resolution light fields is by using the recently developed technique of compressive sensing (CS). The essential element in CS is the codification process, which can be realized by lithographic masks. In this work, a compressive sensing light field camera model based on codification by mask patterns is presented. This research tries to recover a high-resolution light field from a single 2D projection. In the model, the redundant information in the different views of a light field is exploited by the modulation through mask patterns and then an ill-posed underdetermined problem is solved by sparse coding techniques. In order to use the sparse techniques, it is assumed a dictionary where the natural images are compressible enough. Simulations of a dictionary-learning are first performed using a complete dictionary with 5000 light field samples to find the sparsest representation of the light field in order to reconstruct the 4D light field. For the sparse reconstruction, a light field with a 5x5 views is obtained from a single coded 2D projection. The 2D projection is divided into overlapping patches, and then each 2D patch is reconstructed and merged with a median filter to finally obtain the 4D light field. The results indicate the high PSNR (Peak Signal to Noise Ratio) of the reconstructed 4D images.