Photon Limited Non-Blind Deblurring Using Algorithm Unrolling

Abstract

Image deblurring in photon-limited conditions is ubiquitous in a variety of low-light applications such as photography, microscopy and astronomy. However, the presence of photon shot noise due to low-illumination and/or short exposure makes the deblurring task substantially more challenging than the conventional deblurring problems. In this paper we present an algorithm unrolling approach for the photon-limited deblurring problem by unrolling a Plug-and-Play algorithm for a fixed number of iterations. By introducing a three-operator splitting formation of the Plug-and-Play framework, we obtain a series of differentiable steps which allows the fixed iteration unrolled network to be trained end-to-end. The proposed algorithm demonstrates significantly better image recovery compared to existing state-of-the-art deblurring approaches. We also present a new photon-limited deblurring dataset for evaluating the performance of algorithms.

Photon-Limited Blind Deconvolution using Unsupervised Iterative Kernel Estimation

“…The results of the comparison are summarized in Table I. Note that in Table I, we also compare the blind deconvolution methods with the non-blind deconvolution network P4IP [11]. While this network is used iteratively in our scheme as F (.…”

“…Since the blur kernel is not available for this dataset, unlike Levin, the non-blind deconvolution methods i.e. Poisson-PnP [25], PURE-LET [23], P4IP [11] cannot be evaluated on this dataset.…”

“…There are many non-blind Poisson deconvolution solvers. This paper uses a recent method by Sanghvi et al [11]. In this approach, the minimization is solved via an algorithm unrolling that unrolls the plug-andplay alternating direction method of multiplier (PnP-ADMM) into a chain of repeated blocks of steps.…”

“…and noisy-blurred to blur-only denoiser G(.). We use the pre-trained non-blind solver provided by authors of P4IP [11] and use a similar training process for the denoiser G(.). Specifically, we use 2650 images from the Flickr2K dataset [47] as the ground truth.…”

“…This gap is further widened in heavy-noise conditions. For Poisson distributions, there are currently very few deep-learning methods that explicitly model them in the inversion process [11], [12]. (iii) Iterative methods are needed.…”