Unni, VS and Chaudhury, KN (2018) Non-Local Patch-Based Regularization for Image Restoration. In: 25th IEEE International Conference on Image Processing, ICIP 2018, 7-10 October 2018, Megaron Athens International Conference Centre (MAICC)Athens; Greece, pp. 1108-1112.
![]() |
PDF
Int_Con_Img_Pro_1108_2018.pdf - Published Version Restricted to Registered users only Download (1MB) | Request a copy |
Abstract
Several patch-based models have been proposed for image restoration in the literature. A common feature with these models is that patches are used for filtering or optimization, where the aggregation is often performed over a non-local (NL) neighborhood. We propose a NL patch-based regularizer, where patches are used (1) for computing weights between NL neighbors, and (2) for defining a TV-type norm in patch space. A general form of the latter construction was originally proposed by Peyre et al. and later studied by other authors. In most of these proposals, both the weights and the image are treated as variables, which makes the model non-convex. In particular, the corresponding numerical solvers cannot guarantee local optimality. On the other hand, our regularizer is convex. Along with an ell-2 data fidelity term, we apply the regularizer for denoising, deblurring and super-resolution, and develop an efficient ADMM solver for computing the global minimum. An interesting finding is that, while our model is weaker than the non-convex counterparts, the minimizer of the former is generally closer to the ground truth than the reconstruction from the latter. Moreover, we demonstrate that our regularizer can outperform existing regularization techniques for deblurring and super-resolution. © 2018 IEEE.
Item Type: | Conference Proceedings |
---|---|
Publication: | Proceedings - International Conference on Image Processing, ICIP |
Publisher: | IEEE Computer Society |
Additional Information: | Copy right for this article belongs to IEEE Computer Society |
Keywords: | Global optimization; Optical resolving power; Restoration, ADMM; Deblurring; Patch; Regularization; Super resolution, Image reconstruction |
Department/Centre: | Division of Electrical Sciences > Electrical Engineering |
Date Deposited: | 16 Apr 2019 06:48 |
Last Modified: | 16 Apr 2019 06:48 |
URI: | http://eprints.iisc.ac.in/id/eprint/62120 |
Actions (login required)
![]() |
View Item |