SKELLAM DISTRIBUTION BASED ADAPTIVE TWO-STAGE NON-LOCAL METHODS
FOR PHOTON-LIMITED POISSON NOISY IMAGE RECONSTRUCTION
Lingyan Zhao, Jun Zhang
*
and Zhihui Wei
School of Science,
Nanjing University of Science and Technology,
Nanjing, P.R.China, 210094
*
Corresponding author: phil_zj@njust.edu.cn
ABSTRACT
Two-stage non-local methods represented by the Poisson
non-local means (PNLM) method [1] and the non-local
principal component analysis (NLPCA) [2] method perform
well for photon-limited Poisson image reconstruction, where
patch-similarity computation in the non-local reconstruction
stage is guided and affected by a pre-reconstructed image
obtained at the first stage. In this paper, we propose a new
method to provide a better pre-reconstructed image with low
computational cost. Firstly, we propose an adaptive method
for fitting the linear relationship between image intensity and
Poisson parameter; Secondly, we obtain an initial estimated
image according to this relationship. Lastly, we obtain new
pre-reconstructed image by adjusting the initial estimation
according to Skellam distribution. Numerical experiments
show that our method can provide a better pre-reconstructed
image, and therefore can improve the performance of the
PNLM method and reduce the computational cost of the
NLPCA method efficiently.
Index Terms—Poisson image, Skellam distribution, two-
stage strategy, pre-reconstructed image, non-local method
1. INTRODUCTION
Reconstruction of photon-limited Poisson image is urgent
demand and particularly challenging in many application
fields such as security monitoring, astronomy and medical
imaging. Compared with widely used traditional methods
such as variance-stabilizing transformation based methods
[3-4] and empirical Bayesian based regularization methods
[5-6], two-stage non-local methods represented by the
PNLM method and the NLPCA method perform better for
the photon-limited Poisson image reconstruction.
As we all know, accurate computation of similarities
between non-local patches is crucial for any non-local
method. However, information loss and structural damage in
photon-limited Poisson noisy image lead to great difficulties
for accurate computation of the non-local similarities. To
remedy this, both of the PNLM method and the NLPCA
method utilize a two-stage strategy as follows: at the first
stage, a pre-reconstructed image is obtained; at the second
stage, non-local method is used for image reconstruction,
where the computation of the similarities between non-local
patches is guided and affected by the obtained pre-
reconstructed image. The PNLM method utilizes simple
averaging filter at the first stage, and therefore some
important structures such as edges will be blurred which
leads to inaccurate similarity computation. The NLPCA
method combines Poisson distribution based PCA and
sparse Poisson intensity estimation method in a non-local
estimation framework to obtain a better pre-estimated image,
but the computational cost is very high.
From the perspective of statistics, reconstruction of
Poisson image aims to estimate the parameters such as mean
and variance of Poisson distribution. It has been shown that
the intensity of pixels are linearly related to these Poisson
parameters [7] which has been used for Poisson noise
removal in [8]. Motivated by this property, we propose to
obtain a pre-reconstructed image according to this linear
relationship. However, this linear relationship should be
estimated by using the pixels in homogeneous patches. In [7-
8], homogeneous patches are chosen manually which causes
difficulties in practice.
In this paper, we aim to provide a better pre-
reconstructed image which can give better guidance for the
similarities computation at the second reconstruction stage.
Firstly, we propose an adaptive method for fitting the linear
relationship between the intensity of pixels and Poisson
parameters by choosing the homogeneous patches adaptively.
Secondly, we obtain an initial estimated image based on this
relationship according to the intensity of the observed image.
Lastly, we obtain a pre-reconstructed image by adjusting the
initial estimation obtained according to the Skellam
distribution based acceptance range of intensity difference.
The rest of this paper is organized as follows: In
Section2, we describe the PNLM method and the NLPCA
method briefly; In Section3, we introduce our adaptive
method in detail; Numerical experiments are given in
Section4 to demonstrate the efficiency of our method, and
finally, the discussion and conclusion are provided Section 5.
2433978-1-5090-2175-8/17/$31.00 ©2017 IEEE ICIP 2017