The l0-norm-based Blind Image Deconvolution: Comparison and Inspiration

Single image blind deblurring has been intensively studied since Fergus et al.’s variational Bayes method in 2006. It is now commonly believed that the blurkernel estimation accuracy is highly dependent on the pursed salient edge information from the blurred image, which stimulates numerous l0-approximating blind deblurring methods via kinds of techniques and tricks. This paper, however, focuses on the four recent daring attempts which are all based on the simple and direct l0-norm. A systematic comparative analysis is made towards those methods, clarifying their similarities and differences, and providing a benchmark evaluation on both the deblurring quality and computational efficiency. Results have demonstrated that the l0-norm alone is far enough to achieve top blind deblurring performance. Instead, details are to be paid with fairly more attention as working on the problem formulation as well as the algorithmic deduction. Inspired by the success of the bi-l0-l2-norm regularization, an attempt has been made to boost a recently proposed normalized sparsity-based blind deblurring method via simply borrowing core ideas behind the bi-l0-l2-norm regularization. Experimental results show that the boosting approach has leaded to a significant improvement in terms of both accuracy and efficiency. Finally, several possible extensions are discussed towards the bi-l0-l2-norm regularization. Keywords-blind deblurring; camera shake removal; variational Bayes; l0-norm minimization; split Bregman; halfquadratic