Документ взят из кэша поисковой машины. Адрес
оригинального документа
: http://www.adass.org/adass/proceedings/adass03/O2-1/
Дата изменения: Tue Aug 17 02:09:14 2004 Дата индексирования: Tue Oct 2 04:14:49 2012 Кодировка: Поисковые слова: titan |
Superresolution is a useful technique in a variety of applications (Schultz & Stevenson, 1996; Hardie, Barnard, & Armstrong, 1997), and recently, researchers have begun to investigate the use of wavelets for superresolution image reconstruction (Nguyen, Milanfar, & Golub, 2001). We present a new method for superresolution image reconstruction based on the wavelet transform in the presence of Gaussian noise and on an analogous multiscale approach in the presence of Poisson noise. To construct the superresolution image, we use an approach based on maximum penalized likelihood estimation. The reconstructed image is the argument (in our case the superresolution image) that maximizes the sum of a log-likelihood function and a penalizing function. The penalizing function can be specified by an ad hoc smoothness measure, a Bayesian prior distribution for the image (Hebert & Leahy, 1989; Green, 1990), or a complexity measure (Liu & Moulin, 2001). Smoothness measures include simple quadratic functions that measure the similarity between the intensity values of neighboring pixels, as well as non-quadratic measures that better preserve edges. Similar penalty functions result from Markov Random Field (MRF) priors. Complexity measures are usually associated with an expansion of the intensity image with respect to a set of basis functions (e.g. Fourier or wavelet) and count the number of terms retained in a truncated or pruned series (Saito, 1994; Krim & Schick, 1999); the more terms (basis functions) used to represent the image, the higher the complexity measure. Many algorithms (e.g. Expectation-Maximization algorithms, the Richardson Lucy algorithm, or close relatives) have been developed to compute MPLEs under various observation models and penalization schemes.
Wavelets and multiresolution analysis are especially well-suited for astronomical image processing because they are adept at providing accurate, sparse representations of images consisting of smooth regions with isolated abrupt changes or singularities ( e.g. stars against a dark sky). Many investigators have considered the use of wavelet representations for image denoising, deblurring, and image reconstruction; for examples, see Mallat, 1998, and Starck, Murtagh, & Bijaoui, 1998. The proposed approach uses an EM algorithm for superresolution image reconstruction based on a penalized likelihood formulated in the wavelet domain. Regularization is achieved by promoting a reconstruction with low-complexity, expressed in terms of the wavelet coefficients, taking advantage of the well known sparsity of wavelet representations.
The EM algorithm proposed here extends the work of Nowak & Kolaczyk, 2000, and
Figueiredo & Nowak, 2002, which addressed image deconvolution with a method
that combines the efficient image representation offered by the discrete wavelet
transform (DWT) with the diagonalization of the convolution operator obtained in
the Fourier domain. The algorithm alternates between an E-step
based on the fast Fourier transform (FFT) and a DWT-based M-step, resulting in
an efficient iterative process requiring operations per iteration,
where
is the number of pixels in the superresolution image.
From the formulation above, it is clear that superresolution image
reconstruction is a type of inverse problem in which the operator to be
inverted, , is partially unknown due to the unknown shifts and rotations of
the observations. The first step of our approach is to estimate these parameters
by registering the low-resolution observations to one another. Using these
estimates, we reconstruct an initial superresolution image estimate
. This
estimate is used in the third step, where we re-estimate the shift and rotation
parameters by registering each of the low resolution observations to the initial
superresolution estimate. Finally, we use a wavelet-based EM algorithm to solve
for
using the registration parameter estimates. We begin by describing a
wavelet-based method for the Gaussian noise model, and follow that by a
discussion of a multiscale technique for Poisson data. Each of these steps is
detailed below.
The first step in the proposed method is to register the observed low-resolution
images to one another using a Taylor series expansion. This was proposed by
Irani and Peleg, 1991. In particular, let and
be the continuous
images underlying the sampled images
and
, respectively. If
is equal to a shifted, rotated version of
, then we have the
relation
|
After the registration parameters have been initially estimated using the above
method, we use these estimates to calculate an initial superresolution image as
described in Section 4.. This initial image estimate is then used to
refine the registration parameter estimates. The method is the same as above,
but instead of registering a low resolution estimate, , to another low
resolution estimate,
, we instead register it to
. The
results of this refinement are displayed in gray in Figure
1. From these plots, it is clear that the Taylor series
based approach can produce highly accurate results. However, in low SNR
scenarios, where confidence in registration parameter estimates may be low, the
estimates can be further refined at each iteration of the proposed EM algorithm,
as discussed in the following sections.
Note that the motion model considered here encompasses only shift and rotation movement. When recording still or relatively still objects distant from the imaging device, this model is sufficient. More sophisticated models are an area of open research.
The Gaussian observation model in (1) can be written with respect
to the DWT coefficients
, where
and
denotes the
inverse DWT operator (Mallat, 1998):
An analogous formulation is possible for the Poisson noise model. In this case,
photon projections can be described statistically as follows. Photons are
emitted (from the emission space) according to a high resolution intensity
. Those photons emitted from location
are detected (in the detection
space) at position
with transition probability
, where
is the
superresolution operator from (2). In such cases, the measured
data are distributed according to
From these formulations of the problem for Gaussian and Poisson data, the EM
algorithm produces a sequence of estimates
by alternately applying two steps: