Efficient computation of backprojection arrays for 3D light field deconvolution
This work addresses a computational bottleneck in 3D imaging for researchers using plenoptic cameras, though it is incremental as it optimizes an existing step without changing the overall paradigm.
The paper tackles the time-consuming computation of backprojection arrays for 3D light field deconvolution by presenting a new algorithm that rearranges array elements, drastically reducing computation times while yielding identical results to existing methods.
Light field deconvolution allows three-dimensional investigations from a single snapshot recording of a plenoptic camera. It is based on a linear image formation model, and iterative volume reconstruction requires to define the backprojection of individual image pixels into object space. This is effectively a reversal of the point spread function (PSF), and backprojection arrays H' can be derived from the shift-variant PSFs H of the optical system, which is a very time consuming step for high resolution cameras. This paper illustrates the common structure of backprojection arrays and the significance of their efficient computation. A new algorithm is presented to determine H' from H, which is based on the distinct relation of the elements' positions within the two multi-dimensional arrays. It permits a pure array re-arrangement, and while results are identical to those from published codes, computation times are drastically reduced. This is shown by benchmarking the new method using various sample PSF arrays against existing algorithms. The paper is complemented by practical hints for the experimental acquisition of light field PSFs in a photographic setup.