Photon-Limited Blind Deconvolution Using Unsupervised Iterative Kernel Estimation

Blind deconvolution is a challenging problem, but in low-light it even more difficult. Existing algorithms, both classical and deep-learning based, are not designed for this condition. When the photon shot noise strong, conventional methods fail because (1) image does have enough signal-to-noise ratio to perform blur estimation; (2) While deep neural networks powerful, many of them do consider forward process. these simultaneously deblur denoise; (3) iterative schemes known be robust frameworks, they seldom considered requires differentiable non-blind solver. This paper addresses above challenges by presenting an unsupervised blind method. At core method reformulation general framework from image-kernel alternating minimization purely kernel-based minimization. leads new scheme that backpropagates unsupervised loss through pre-trained solver update kernel. Experimental results show proposed achieves superior than state-of-the-art algorithms conditions.