Super-Gaussian Fields: A Novel Model to Image Deblurring
Abstract: Blind image deblurring is a challenging problem due to its ill-posed nature, of which the success is closely related to a proper image prior. Although most of sparsity-based priors on the gradient filters have been successfully applied, they are inherently limited by the fact that they only explore local coherence in natural image statistics and thus cannot model more complicated structures. We aim to leverage Markov random fields (MRFs) to break the limitation. Due to the intractable partition function, however, traditional MRFs often learn universal filters for various images, resulting in unsatisfactory performance. Motivated by this, we propose a novel MRF-based image prior, referred to as Super-Gaussian Fields. Specifically, we depict potentials by using super-Gaussian distributions, leading to image-specific filters. Relying on the prior and Bayesian MMSE, we proposed an effective image deblurring method. Theory analyses show that the proposed method can effectively avoid local minimum, and can learn image-adaptive sparsity-promoting filters that highlight image structures for kernel estimation. Most importantly, with the theory support, the proposed method can be extended to various scenarios, e.g., face, text, and low-illumination image deblurring. Extensive experiments demonstrate the theoretical advantages and practical effectiveness of the proposed method.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Xiaochun_Cao3
Submission Number: 3800
Loading