Characterization of Plenoptic Imaging Systems and Efficient Volumetric Estimation From Plenoptic Data

Light-field cameras in conjunction with computational refocusing can be used to produce volumetric estimates of an imaged scene. However, these estimates are often dominated by image blur in the depth direction from objects not in each synthesized focal plane. Tomographic algorithms have been shown to be effective in creating volumetric estimates from plenoptic data but are often prohibitively slow. Deconvolution would be an attractive solution due to processing speed, but existing image synthesis equations are shift-variant. This research proposes an alternate refocusing transformation that makes the core problem described in continuous coordinates shift-invariant so that deconvolution is a viable solution. Shift-invariance of the new refocusing transform is demonstrated mathematically. Furthermore, the discretization involved in the imaging system and refocusing algorithm are characterized with respect to shift-variance in order to identify potential sources of artifacts and to propose potential mitigating steps where possible. While the sampled light field is not directly invertible, experimental data are used to demonstrate that regularized deconvolution using the derived synthesis equations produces improved results compared to the base focal stack in both synthetic examples and actual camera data.