Image restoration based on neural networks and nonlocal denoising
Închide
Articolul precedent
Articolul urmator
763 10
Ultima descărcare din IBN:
2023-03-24 09:50
Căutarea după subiecte
similare conform CZU
004.932+519.6+519.8 (1)
Informatică aplicată. Tehnici bazate pe calculator cu aplicații practice (438)
Matematică computațională. Analiză numerică. Programarea calculatoarelor (123)
Cercetări operaționale (OR) teorii şi metode matematice (167)
SM ISO690:2012
CRUZ, C., KATKOVNIK, Vladimir, EGIAZARIAN, Karen O.. Image restoration based on neural networks and nonlocal denoising. In: Materials Science and Condensed Matter Physics, Ed. 9, 25-28 septembrie 2018, Chișinău. Chișinău, Republica Moldova: Institutul de Fizică Aplicată, 2018, Ediția 9, p. 285.
EXPORT metadate:
Google Scholar
Crossref
CERIF

DataCite
Dublin Core
Materials Science and Condensed Matter Physics
Ediția 9, 2018
Conferința "International Conference on Materials Science and Condensed Matter Physics"
9, Chișinău, Moldova, 25-28 septembrie 2018

Image restoration based on neural networks and nonlocal denoising

CZU: 004.932+519.6+519.8

Pag. 285-285

Cruz C.1, Katkovnik Vladimir21, Egiazarian Karen O.12
 
1 Noiseless Imaging Oy, Tampere,
2 University of Tampere
 
Disponibil în IBN: 12 februarie 2019


Rezumat

Nonlocal image restoration based on patch grouping and collaborative filtering has gained popularity in the recent decade. The BM3D frame [1] is an example of such a technique; it utilizes both local sparsity of small image patches and the group-sparsity of collections of self-similar image patches. As sparsifying transforms in the spatial and in the similarity domains, both fixed transforms (e.g. DCT, wavelets) and data-adaptive transforms (e.g. SVD, PCA) have been used. In [2], an alternating application of specially designed nonlocal collaborative filters in an iterative image restoration scheme is proposed. This work is in line with the current trends in image processing, where the accurate formulation of the prior can be omitted in favor of a good denoising algorithm embedded in the iterations. We prove that the used priors result in the efficient filters and that the alternating application of these filters leads to exceptional results. Recently, image restoration received a new boost of interest through the application of advanced machine-learning methods, particularly deep convolutional neural networks (CNNs) [3]. The main advantages of CNN-based filters lie in their ability to learn and extract complex image features, and, even more importantly, in their efficient implementation on graphics processing units (GPUs). However, the application of CNNs in the problem of image restoration comes with certain drawbacks. First, learning is often a very time consuming process, which can take from a few hours to several days. Furthermore, CNNs are inferior to nonlocal methods when it comes to images exhibiting a high degree of self-similarity (e.g. regular textures, edges). In our recent paper [4], we introduced a novel paradigm called NN3D that combines the advantages of CNN-based and nonlocal filters through a simple iterative modular framework, achieving state-of-the-art results in image denoising. In this paper, we propose a novel approach to image restoration based on a combination of the ideas from [2] and [4], which consists of applying NN3D as a new denoising prior in the ‗plug-and-play‘ framework of image restoration. Extensive simulations show that the developed approach leads to state-of-the-art performance.