Compressive total variation for image reconstruction and restoration☆
Introduction
Many image processing problems can be formulated as an inverse problem, in which the data is assumed to be obtained approximately by applying a linear operator on an image with the additive noise . In most of the cases, solving can be via However, this problem is ill-posed in the sense that directly inverting would lead to bad and possibly multiple solutions. It is necessary and even desirable to constrain the solutions through a regularization, which provides the prior knowledge of images that one wants to reconstruct or recover. The regularization is usually used to avoid non-uniqueness of solutions and improve the quality of solutions. A general model for such an inverse problem is where is the data fitting term, and is the regularization term (or penalty term), and is a positive parameter to balance two terms, and is an optimal solution of the model.
Usually, the data fitting term is taken as . And in this subsection, we only give a simple review of the regularization term .
A classical regularization is the total variation (TV) as follows where denote the horizontal and vertical partial derivative operators (with certain boundary conditions assumed), respectively, and is the gradient operator in the discrete setting. Here we give the mathematical definition only in the discrete setting. It is originated by Rudin-Osher-Fatemi [1], which is referred to as the ROF model. It is widely used in image processing applications, such as denoising [1], deconvolution [2], MRI reconstruction [3], recognition [4], inpainting [5], and super-resolution [6] and so on. This TV model is isotropic, and later an anisotropic formulation was addressed in the literature [7] (see also [8] and other references). For any two-dimensional (2D) image, the anisotropic TV is defined by
There exist some different penalties as alternatives to TV. A few notable examples of nonconvex replacements of TV are quasi-norm of for [9], [10], norm of for [11], [12], [13]. And there also exists higher order TV-total generalized variation (TGV) (see Section 2.2) [14], [15], which is more precise in describing intensity variations in smooth regions, and thus reduces oil painting artifacts while still being able to preserve sharp edges like TV does. Except these TV and its variants, there also exists some combination of TV (or TGV) and different waves, for example TV+wavelet [16], TGV+shearlet [17], [18] and so on. The connection between TV (or even TGV) and wavelet frames also has been analyzed in [19].
Although TV regularization can preserve edge information, it does not take full advantage of the similarities in images. More regularizations, for example the sparse regularization based dictionary learning and the low-rank regularization based on similar image patches, have been introduced in [20], [21], [22] and [23], [24], respectively. In here, we do not recall them in detail.
In this subsection, we delineate the motivation. First, we recall the interpretation of TV from the perspective of compressive sensing [25], [26]. The compressive sensing aims to reconstruct a signal or an image from an underdetermined system of linear equations , provided that the signal or image is sufficiently sparse or sparse in a transform domain. For instance, an image is mostly sparse after taking the gradient transform. Mathematically, this problem is equal to minimize the norm of the image gradient, i.e., . However, such method is NP-hard and thus computationally infeasible in high dimensional settings. Candès et al. [25] replace by a convex relaxation , i.e., the norm of , which is the TV. They showed that images could be reconstructed via norm through numerical tests.
Now, we show the motivation.
- (i)
Low-rank Matrix Completion: Matrix completion has been a valuable but difficult task in image inpainting and recommender systems. Given an image with missing parts or corrupted regions, the goal is to find the missing pixel values while ensuring the completed results visually reasonable. For an image, it is probably of (approximately) low rank structure [27]. Candès et al. [28] first solved the matrix completion problem by approximating the rank with nuclear norm as follows where is a projection operator on index set , i.e., for , and for . More works about low-rank matrix completion , readers can refer to an overview [29].
- (ii)
Total Nuclear Variation (TNV): For a separable image , all channels at any point should share a common gradient direction. For example, color images using three channels (red, green, blue), share a common gradient direction. Furthermore, having shared directions is equivalent to having a rank-one Jacobian. Therefore Holt [30] thus proposed using nuclear norm in the Jacobian framework, which was called total nuclear variation.
- (iii)
Low-rank Prior: The similar image patches have similar underlying structures. Thus the matrix constructed from stacking the similar patches together has low rank. Those successful works in [23], [24] show that the methods based on low rank minimization can capture the underlying multi-scale structure and provide good representation for images.
- (iv)
Robust Principal Component Analysis (RPCA): Some 2D data, has the superposition of a low-rank component and a sparse component, for example video surveillance, face recognition, latent semantic indexing and ranking and collaborative filtering. In [31], [32], the authors proposed a very convenient convex program called robust principal component analysis, to recover both the low-rank component and the sparse component as follows where is the balancing parameter and is the observations. More works about RPCA, readers can refer to [33], [34].
- (v)
Compressive Phase Retrieval (CPR): In order to solve sparse phase retrieval problem, Ohlsson et al. [35] proposed a model via the lifting technique, which involves a symmetric positive semi-definite, sparse and low-rank matrix , as follows where is the observations, with or , and . This model is called compressive phase retrieval via the lifting (CPRL) in [35]. Since then, compressive Phase Retrieval has been studied by many scholars [36], [37], [38]. And this model has been extended to compressive affine phase retrieval in [39].
Note that gradient matrices of an image are sparse. And the image also has (approximately) low rank structure. Motivated by robust principal component analysis and compressive phase retrieval, we conjecture that these gradient matrices also are (approximately) low-rank.
The main contributions of this paper are three folds.
- (1)
We consider that the gradient matrices of an image are not only sparse but also (approximately) low-rank, and verify it by numerical tests (Section 2.1) and theoretical analysis (Section 2.2). In order to characterize the low-rank prior information, we introduce the Nuclear norm Total (Generalized) Variation (NT(G)V) (Section 2.3). And we establish a model-compressive total variation (CTV), which reflects these prior informations of the image (Section 2.4).
- (2)
Based on inertial proximal ADMM, we design a concrete algorithm to solve our model (Section 3).
- (3)
To demonstrate the effectiveness of the proposed method, we test many numerical examples, including MRI reconstruction from incomplete data (Section 4.1), denoising from noisy data (Section 4.2), and deblurring from blurring data (Section 4.3). The tests show that proposed CTV method is better than that of classic TV, TGV, Shearlet-TGV, TV and BM3D methods in testing for piecewise constant images. Our method can not only hold sharper edges but also preserve various image features. And proposed method is comparable to TGV, Shearlet-TGV, and TV methods in testing for natural images.
Our notation is standard, as used above in this section. The standard inner product is denoted by . Let denote the norm (or quasi-norm) of matrix or vector, i.e., or . For any matrix with rank , let be the singular value decomposition (SVD) of , where is the singular vector with . We denote as the nuclear norm of the matrix . The superscript denotes the matrix/vector transpose operator. And we denote by identity matrix, by zero matrix and denotes the set of symmetric matrices. In all of this paper, we use boldfaced letter to denote matrix or vector.
The rest of the paper is organized as follows. In Section 2, we test the rank of gradient matrices of images and find that gradient transform matrices are not only sparse but also low-rank or approximately/relatively low-rank, and establish a model-compressive total variation (CTV), which reflects these priors of the image. In Section 3, we design a concrete algorithm based inertial proximal ADMM with provable convergence to solve our model. In Section 4, we compare our CTV method with some other existing methods in MRI reconstruction, image denoising and image deblurring. Conclusions and discussions are given in Section 5.
Section snippets
Compressive total variation model
In this section, we establish our model. First, we test the rank of gradient matrices of images to verify that gradient matrices are indeed low-rank (or approximately/relative low-rank).
Numerical algorithm for CTV
In this section, we design a concrete algorithm to solve the CTV model (2.7).
In fact, by splitting in the sparse term, in the low-rank term and in the low-rank term, one has Let and and , then optimization problem (3.1) can be rewritten as which is a
Numerical examples of CTV
In this section, we present numerical experiments in compressible images to demonstrate the efficiency of our model (2.7).
We test on both simulated data and real in vivo data. Several sets include incomplete spectral data (DFT): 405 × 405 brain magnetic resonance images, 512 × 512 foot magnetic resonance images; noisy data: 200 × 200 Circles image, 256 × 256 House image; blurring data: 256 × 256 Binary image, 256 × 256 Cameraman image, 256 × 256 Text image, 512 × 512 Lena image. For better
Conclusions and discussion
In this paper, based on the facts that the image is (approximately) low-rank and the corresponding gradient transform matrices are sparse, we consider that the gradient matrices and are not only sparse but also (approximately) low-rank. We verify this conclusion by numerical tests and theoretical analysis. Based on these prior knowledge, we propose the compressive total variation (CTV) model for image processing applications. Our model inherits the superior performance of TGV,
CRediT authorship contribution statement
Peng Li: Resources, Project Administration, Supervision, Funding Acquisition, Conceptualization, Investigation, Validation, Formal Analysis, Methodology, Data Curation, Software, Visualization, Writing - original draft, Writing - review & editing. Wengu Chen: Funding Acquisition, Conceptualization, Writing - review & editing. Michael K. Ng: Funding Acquisition, Conceptualization, Methodology, Validation, Formal Analysis, Data Curation, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors thank Professors Raymond Honfu Chan and Jinshan Zeng, and Drs. Chao Zeng and Taixiang Jiang for their help in the preparation of this paper.
References (57)
- et al.
Nonlinear total variation based noise removal algorithms
Physica D
(1992) - et al.
Selesnick improved sparse low-rank matrix estimation
Signal Process.
(2017) - et al.
Total variation blind deconvolution
IEEE Trans. Image Process.
(1998) - et al.
Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information
IEEE Trans. Inform. Theory
(2006) - et al.
Total variation models for variable lighting face recognition
IEEE Trans. Pattern Anal. Mach. Intell.
(2006) - et al.
Mathematical models for local nontexture inpainting
SIAM J. Appl. Math.
(2002) - et al.
Image super-resolution by TV-regularization and bregman iteration
J. Sci. Comput.
(2008) - et al.
Decomposition of images by the anisotropic Rudin-Osher-Fatemi model
Comm. Pure Appl. Math.
(2003) - et al.
Anisotropic total variation regularized -approximation and denoising/deblurring of 2D bar codes
Inverse Probl. Imaging
(2011) Exact reconstruction of sparse signals via nonconvex minimization
IEEE Trans. Signal Process.
(2007)
A weighted difference of anisotropic and isotropic total variation model for image processing
SIAM J. Imaging Sci.
Truncated models for sparse recovery and rank minimization
SIAM J. Imaging Sci.
Minimization methods for signal and image reconstruction with impulsive noise removal
Inverse Problems
Total generalized variation
SIAM J. Imaging Sci.
Second order total generalized variation (TGV) for MRI
Magnet. Resonance Med.
A fast alternating direction method for TVL1-L2 signal reconstruction from partial Fourier data
IEEE J. Sel. Top. Signal Process.
A new detail-preserving regularization scheme
SIAM J. Imaging Sci.
Shearlet-TGV Based Fluorescence Microscopy Image DeconvolutionCAM Report
Image restoration: Total variation, wavelet frames, and beyond
J. Amer. Math. Soc.
Emergence of simple-cell receptive field properties by learning a sparse code for natural images
Nature
Image deblurring via total variation based structured sparse model selection
J. Sci. Comput.
Low rank prior and total variation regularization for image deblurring
J. Sci. Comput.
Stable signal recovery from incomplete and inaccurate measurements
Comm. Pure Appl. Math.
Compressed sensing
IEEE Trans. Inform. Theory
Fast and accurate matrix completion via truncated nuclear norm regularization
IEEE Trans. Pattern Anal. Mach. Intell.
Cited by (15)
Digital inpainting of mural images based on DC-CycleGAN
2023, Heritage ScienceLatent-space Unfolding for MRI Reconstruction
2023, MM 2023 - Proceedings of the 31st ACM International Conference on MultimediaMural inpainting progressive generative adversarial networks based on structure guided
2023, Beijing Hangkong Hangtian Daxue Xuebao/Journal of Beijing University of Aeronautics and Astronautics