Light Field Blind Deconvolution

We address for the first time the issue of motion blur in light field images captured from plenoptic cameras (instead of camera arrays), where the spatial sampling in each view is decimated. We propose a solution to the estimation of a sharp light field given a blurry one, when the motion blur point spread function is unknown, i.e., the so-called blind deconvolution problem. Unfortunately, the (decimated) spatial sampling of each light field view does not allow the use of current blind deconvolution approaches for traditional cameras. Also, due to the complexity of the imaging model, we investigate first the case of uniform (shift-invariant) blur of Lambertian objects, i.e., when objects are sufficiently far away from the camera to be approximately invariant to depth changes and their reflectance does not vary with the viewing direction. We introduce a highly parallelizable model for light field motion blur that is computationally and memory efficient. We then adapt a regularized blind deconvolution approach to our model and demonstrate its performance on both synthetic and real light field data.