Single Image Super-Resolution Based on Nonlocal Sparse and Low-Rank Regularization
Image super resolution (SR) is an active research topic to obtain an high resolution (HR) image from the low resolution (LR) observation. Many results of existing methods may be corrupted by some artifacts. In this paper, we propose an SR reconstruction m
- PDF / 1,990,892 Bytes
- 11 Pages / 439.37 x 666.142 pts Page_size
- 95 Downloads / 194 Views
Abstract. Image super resolution (SR) is an active research topic to obtain an high resolution (HR) image from the low resolution (LR) observation. Many results of existing methods may be corrupted by some artifacts. In this paper, we propose an SR reconstruction method for single image based on nonlocal sparse and low-rank regularization. We form a matrix for each patch with its vectorized similar patches to utilize the redundancy of similar patches in natural images. This matrix can be decomposed as the low rank component and sparse part, where the low rank component depictures the similarity and the sparse part depictures the fine differences and outliers. The SR result is achieved by the iterative method and corroborated by experimental results, showing that our method outperforms other prevalent methods.
Keywords: Super resolution self-similarity
1
·
Low-rank
·
Sparsity
·
Nonlocal
Introduction
High resolution (HR) images, compared with low resolution (LR) images, provide more details. However, it is often difficult to acquire HR images due to limitations such as digital imaging systems or imaging environments [1]. Therefore, image super resolution (SR) reconstruction, a technique to reconstruct the HR image from one or several LR images, has become an active topic in many fields such as image processing and computer vision. For single image SR reconstruction, the simplest type of method is the interpolation-based method, such as bilinear, bicubic and other resampling methods [2,3], in which an interpolation kernel was applied to estimate the missing pixels in the HR image. This tends to produce blur edges with ringing and jagged artifacts [1]. In this regards, many structure priors and nonlocal methods were proposed as regularizers to improve results. One popular prior is the total variation (TV) based method [4], in which the 1 norm of the image derivative were applied. c Springer International Publishing Switzerland 2016 R. Booth and M.-L. Zhang (Eds.): PRICAI 2016, LNAI 9810, pp. 251–261, 2016. DOI: 10.1007/978-3-319-42911-3 21
252
C. Liu et al.
In recent years, learning based methods have become prevalent for image SR. This type of methods recovers lost high frequency details with the help of the trained LR and HR image patch pairs. For instance, [5] proposed an example based method, and the relationship between the patch pairs is estimated by the Markov random field. Inspired by the locally linear embedding (LLE) methods from manifold learning [6], the assumption that the LR and HR local patch pairs have similar local geometries in two distinct feature spaces was established [7]. Then, the local geometry from the LR patch and its k-neighbors was learned, and mapped to the HR space to reconstruct the SR image. [8,9] proposed the sparse coding based SR methods, which assumed that the sparse representation is consistent between the LR and HR patch pairs. Later, the nonlocal self-similarity [10], a property for nature images, was introduced into the image processing. [1,11] extended the nonlocal me
Data Loading...
