Block-based streaming blind deconvolution for space-variant turbulence mitigation

Abstract

A block-based, streaming, multi-frame blind deconvolution (MFBD) mitigation method is presented for restoration of imagery impacted by space-variant atmospheric turbulence. An incremental approach is taken, referred to as “streaming”, which operates on each new frame as it arrives. For each new frame, an optimization is initiated to minimize the error between the new frame and forward modeling of the object and point spread function (PSF). To adapt to space-variant turbulence conditions, the fundamental units of operation are small, overlapping blocks of pixels extracted from the entire frame. The optimization for each new block is seeded using the solution from the same block area in the previous frame to produce a cumulative, improved, block solution. For each block, the algorithm implements stochastic gradient descent, and alternates between seeking a solution for the PSF and the object, while holding the other constant. To assist in regularizing the PSF solutions during the processing, the PSF estimate is projected onto a sparse dictionary representation of PSFs. The block solutions are combined using an overlap/add method to produce the object estimate after each frame. Our method can utilize either a Picard iterative process or a limited memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm with bound constraints. A comparison of results is presented using data simulated with high-fidelity to real systems and environments. While incremental methods for blind deconvolution have been reported in the literature, our method has been developed explicitly for the routine mechanics of blind deconvolution.