Image denoising with patch estimation and low patch-rank regularization
Bo Li, Hongyi Wang
College of Mathematics and Information Science , Nanchang Hangkong University
Nanchang, China
National Engineering Research Center of Digital Life, Sun Yat-sen University
Email: bolimath@gmail.com
Abstract—In this paper, we propose an image denoising
algorithm for data contaminated by Poisson noise using patch
estimation and low patch-rank regularization. In order to
form the data fidelity term, we take the patch-based poisson
likelihood, which will effectively remove the ’blurring’ effect.
For the sparse prior, we use the low patch-rank as the
regularization, avoiding the choosing of dictionary. Putting
together the data fidelity and the prior terms, the denoising
problem is formulated as the minimization of a maximum a
posteriori (MAP) objective functional involving three terms: the
data fidelity term; a sparsity prior term, in the form of a low
patch-rank regularization ;and a non-negativity constraint (as
Poisson data are positive by definition). Experimental results
show that this algorithm achieved better results via giving
specific constraints on different component and get faster
convergence rate .
Keywords-Patch estimation; Low patch-rank; Proximal split-
ting method
I. INTRODUCTION
An important task in mathematical image processing is
image denoising. The general idea is to regard a noisy image
f as being obtained by corrupting a noiseless image u;given
a model for the noise corruption, the desired image u is a
solution of the corresponding inverse problem. Some image
denoising algorithms assume that the noise is normally dis-
tributed and additive and many algorithms has been proposed
for reconstructing u from f . Since the inverse problem is
generally illposed, most denoising procedures employ some
sort of regularization. A very successful algorithm is that of
Rudin, Osher, and Fatemi [1,2], which uses total-variation
regularization. The ROF model regards u as the solution to
a variational problem, to minimize the functional
F (u)=
Ω
|∇u| + λ
Ω
|f − u|
2
where Ω is the image domain and λ is a parameter to be
chosen. The first term is a regularization term, the second
a data-fidelity term. Minimizing F (u) has the effect of
diminishing variation in u, while keeping u close to the
data f . The size of the parameter λ determines the relative
importance of the two terms. Like many denoising models,
the ROF model is most appropriate for signal independent,
additive Gaussian noise. See [3] for an explanation of this
in the context of Bayesian statistics.
Many images, however, contain noise that satisfies a
Poisson distribution. The magnitude of Poisson noise varies
across the image, as it depends on the image intensity.
This makes removing such noise very difficult. A familiar
example is that of radiography. The signal in a radiograph
is determined by photon counting statistics and is often
described as particle-limited, emphasizing the quantized
and non-Gaussian nature of the signal. Removing noise of
this type is a more difficult problem. Besbeas et al. [4]
review and demonstrate wavelet shrinkage methods from the
now classical method of Donoho [5] to Bayesian methods
of Kolaczyk [6] and Timmermann and Novak [7]. These
methods rely on the assumption that the underlying intensity
function is accurately described by relatively few wavelet
expansion coefficients. Kervrann and Trubuil [8] employ an
adaptive windowing approach that assumes locally piecewise
constant intensity of constant noise variance. The method
also performs well at discontinuity preservation. Jonsson,
Huang, and Chan [9] use total variation to regularize positron
emission tomography in the presence of Poisson noise, and
use a fidelity term similar to what we use below. Other
methods, such as sparse-inducing regularizations have also
been proposed in [10,11].
Recently, F.X.Dupe, M.J.Fadili[12] propose an image de-
convolution algorithm under poisson noise, it considers the
image formation model where an input image of n pixels x is
contaminated by Poisson noise. The objective is to minimize
the following functional
x =min
x
− ln L(x)+φ(x)+ι
C
where
The first term − ln L(x) is data-fidelity function, it’s an
appropriate likelihood function, obtained by the distribution
of the observed data y given an original x, and it reflects
the Poisson statistics of the noise. However, because it uses
pixel-by-pixel max-likehood estimation as the data-fidelity
term, the reconstructed image will be ’blurring’.
The second term φ(x) is a regularization for x, as the
image is supposed to be economically (sparsely) represented
in a pre-chosen dictionary D, so the regularization term φ(x)
is a sparsity-inducing penalty . The difficult problem is how
to choose the proper dictionary.
2012 Fourth International Conference on Digital Home
978-0-7695-4899-9/12 $26.00 © 2012 IEEE
DOI 10.1109/ICDH.2012.12
224