SelfDeblur with Sparsity Enforced Bregman Learning

Abstract

Learning-driven methods have revolutionized the field of blind deconvolution, with SelfDeblur as a pioneering method. It uses deep image priors to jointly estimate blur kernels and latent clear images, employing deep neural networks to automatically learn these priors and effectively capture their statistical properties. Building on the success of SelfDeblur, this paper proposes a novel sparse optimization method, named B-SelfDeblur, which employs linearized Bregman iterations to train neural networks. Our approach addresses the challenges of processing large-scale blurry images by starting with minimal parameters and incrementally adding necessary ones in an inverse scale space manner. Additionally, we incorporate a smoothing technique for the total variation (TV) regularization term, which ensures our loss function is continuously differentiable. Experimental results demonstrate that B-SelfDeblur can significantly optimize network weight storage while maintaining deblurring effectiveness. Furthermore, we provide a partial theoretical analysis and ablation studies, which confirm that our loss function is monotonically decreasing and that our method is well-suited for alternating iterative optimization.