Deep Unfolding of a Proximal Interior Point Method for Image Restoration

Variational methods are widely applied to ill-posed inverse problems for they have the ability to embed prior knowledge about the solution. However, the level of performance of these methods significantly depends on a set of parameters, which can be estimated through computationally expensive and time-consuming methods. In contrast, deep learning offers very generic and efficient architectures, at the expense of explainability, since it is often used as a black-box, without any fine control over its output. Deep unfolding provides a convenient approach to combine variational-based and deep learning approaches. Starting from a variational formulation for image restoration, we develop iRestNet, a neural network architecture obtained by unfolding a proximal interior point algorithm. Hard constraints, encoding desirable properties for the restored image, are incorporated into the network thanks to a logarithmic barrier, while the barrier parameter, the stepsize, and the penalization weight are learned by the network. We derive explicit expressions for the gradient of the proximity operator for various choices of constraints, which allows training iRestNet with gradient descent and backpropagation. In addition, we provide theoretical results regarding the stability of the network for a common inverse problem example. Numerical experiments on image deblurring problems show that the proposed approach compares favorably with both state-of-the-art variational and machine learning methods in terms of image quality.

Mots clés

Interior point methods Interior point method proximal algorithms deep unfolding Neural networks regularization neural network image restoration

Domaines

Optimisation et contrôle [math.OC] Apprentissage [cs.LG] Traitement des images [eess.IV]

Fichier principal

DeepUnfolding.pdf (2.76 Mo)

1_gamma.pdf (16.4 Ko)

1_lambda.pdf (18.35 Ko)

1_mu.pdf (20.88 Ko)

35028_gamma.pdf (15.95 Ko)

35028_lambda.pdf (18.43 Ko)

35028_mu.pdf (20.15 Ko)

87015_gamma.pdf (18.96 Ko)

87015_lambda.pdf (19.16 Ko)

87015_mu.pdf (18.24 Ko)

_ircnn_1_906_fcnn_1_856_irestnet_1_909.pdf (215.36 Ko)

_ircnn_87015_685_irestnet_87015_713.pdf (149.57 Ko)

_mlp_35028_860_pdhg_35028_772_ircnn_35028_840_irestnet_35028_883.pdf (243.99 Ko)

cnnpp_ircnn.pdf (81.36 Ko)

gaussian_1_6_std_0008_BSD500_res.pdf (46.97 Ko)

gaussian_1_6_std_001_005_BSD500_res.pdf (47.06 Ko)

gaussian_3_std_004_BSD500_res.pdf (47.3 Ko)

groundtruth_1_blurred_1_576_var_1_844_epll_1_849.pdf (303.38 Ko)

groundtruth_35028_blurred_35028_509_var_35028_833_epll_35028_839.pdf (351.26 Ko)

groundtruth_87015_blurred_87015_344_var_87015_622_epll_87015_553.pdf (456.07 Ko)

iopart10.clo (4.59 Ko)

iopart12.clo (4.59 Ko)

lkmu.pdf (82.11 Ko)

motion3_std_001_BSD500_res.pdf (45.42 Ko)

motion8_std_001_BSD500_res.pdf (50.72 Ko)

net_archi.pdf (67.17 Ko)

prox_barrier_l2norm_2D.pdf (61.61 Ko)

prox_barrier_range.pdf (122.64 Ko)

square_7_std_001_BSD500_res.pdf (41.65 Ko)

Origine	Fichiers produits par l'(les) auteur(s)

Marie-Caroline CORBINEAU : Connectez-vous pour contacter le contributeur

https://hal.science/hal-01943475

Soumis le : mardi 21 janvier 2020-19:31:21

Dernière modification le : lundi 22 avril 2024-17:31:14

Dates et versions

hal-01943475 , version 1 (07-12-2018)

hal-01943475 , version 2 (15-07-2019)

hal-01943475 , version 3 (16-07-2019)

hal-01943475 , version 4 (21-01-2020)

Identifiants

HAL Id : hal-01943475 , version 4
ARXIV : 1812.04276
DOI : 10.1088/1361-6420/ab460a

Citer

Carla Bertocchi, Emilie Chouzenoux, Marie-Caroline Corbineau, Jean-Christophe Pesquet, Marco Prato. Deep Unfolding of a Proximal Interior Point Method for Image Restoration. Inverse Problems, In press, ⟨10.1088/1361-6420/ab460a⟩. ⟨hal-01943475v4⟩

Exporter

BibTeX XML-TEI Dublin Core DC Terms EndNote DataCite

Collections

INRIA CVN CENTRALESUPELEC INRIA2 TDS-MACS UNIV-PARIS-SACLAY GS-ENGINEERING GS-COMPUTER-SCIENCE HUB-IA

639 Consultations

1240 Téléchargements