%%% FO Modif
\documentclass[11pt,twoside]{article} % Leave intact
\usepackage{adassconf}
\begin{document} % Leave intact
\paperID{O2-3}
%%%% ID=O2-3
\title{Scale sensitive deconvolution}
\author{S.\ Bhatnagar and T.J.\ Cornwell}
\affil{National Radio Astronomy Observatory (NRAO), Socorro, NM - 87801\\
Emails: sbhatnag@aoc.nrao.edu, tcornwel@aoc.nrao.edu}
%\altaffiltext{1}{Now at the National Radio Astronomy Observatory (NRAO), Socorro, NM - 87801}
\contact{S. Bhatnagar}
\email{sbhatnag@aoc.nrao.edu}
\paindex{Bhatnagar, S.}
\aindex{Cornwell, T. J.}
\authormark{Bhatnagar \& Cornwell}
\keywords{data analysis: radio interferometry, deconvolution}
%-----------------------------------------------------------------------
% Abstract
%-----------------------------------------------------------------------
\begin{abstract} % Leave intact
Aperture synthesis radio telescopes measure the Fourier tran\-s\-form of
the sky brightness distribution. However the Point Spread Function
(PSF) of such telescopes has significant and widespread side-lobes,
which needs to be deconvolved from the images. Existing deconvolution
algorithms can be thought of as decomposing the image into a set of
delta functions (scale less basis). This uses more degrees of freedom
than necessary and is not optimal for extended emission. In this
paper we present an iterative scale sensitive deconvolution algorithm
for radio interferometric imaging, which attempts to minimize the
degrees of freedom used to represent the signal (spatially correlated
pixels).
%The problem of deconvolution can be thought of as a search algorithm
%for a function P(x) in the image space which, when passed through the
%telescope measurement equation, fits the observed data. Most
%algorithms in use now, gain in speed by setting P(x) to a delta
%function at the location of each pixel in the image (zero correlation
%scale) and search for the amplitude at each pixel. They are therefore
%not optimal when emission is extended and use more degrees of freedom
%than necessary.
%The convolution of the PSF with P(x), which is the dominant
%cost in a search algorithm, in this case reduces to a shift-and-scale
%operation. Such algorithms however search only for the amplitude of
%each pixel in the image and are insensitive to finite correlation
%lengths in the image. These algorithms are therefore not optimal for
%extended emission (emission at a number of spatial scales). They use
%more parameters to represent an extended source than is necessary,
%resulting into deconvolution errors. For example, ideally a two
%dimensional Gaussian source can be represented by 6 parameters (the
%amplitude, variance, position angle and location of the Gaussian)
%rather than with one parameter per pixel (the amplitude for each
%pixel). This results into breaking-up of the source or striping.
%With the increase in telescope sensitivity, such errors can limit the
%achievable dynamic range in the images.
%In this paper we present an iterative scale sensitive deconvolution
%algorithm for radio interferometric imaging, which attempts to minimize
%the degrees of freedom used to represent the signal (spatially
%correlated pixels). We demonstrate that the algorithm is indeed
%sensitive to the local scale in the image. Performance issues and
%comparisons of the results with other successful deconvolution
%algorithms will also be discussed.
\end{abstract}
%-----------------------------------------------------------------------
% Main Body
%-----------------------------------------------------------------------
% Place the text for the main body of the paper here. You should use
% the \section command to label the various sections; use of
% \subsection is optional. Significant words in section titles should
% be capitalized. Sections and subsections will be numbered
% automatically.
\section{Introduction}
A radio interferometric telescope incompletely measures the visibility
function, at discrete points. The Fourier transform of the visibility
function, called the Dirty Image ($I^D$), is the convolution of the
true image ($I^o$) with the telescope Point Spread Function (PSF)
\begin{equation}
\label{DM}
I^D = {\mathcal{FT}}(V^oS) = I^{o}\star B
\end{equation}
where ($B$) is the PSF, $S$ the visibility sampling function and $V^o$
is the observed visibility.
The goal of deconvolution algorithms is to estimate a sky image model
$I^M$, such that the model visibility $V^M={\mathcal{FT}}^{-1}(I^M
\star B)$ fits the observed visibility to the extent allowed by the
noise. A generalized model image $I^M$ can be expressed as a linear
sum of Pixel models
\begin{equation}
\label{O2-3:MODIMG}
I^M=\sum_{k=0}^{N_M} P_m(\vec{p_k})
\end{equation}
where $P(\vec{p_k})$ is the pixel model, $\vec{p}$ are the parameters
for the amplitude, location and the shape of $P$. The problem of
optimal deconvolution then reduces to solving for $I^M$ with a minimal
set $\{\vec{p}\}$ allowed by the data.
The current popular image deconvolution algorithms (Karovska, 2002),
like CLEAN (and its variants (Clark, 1980; Cornwell et al., 1990) and
MEM (and its variants (Cornwell\&Evans, 1985)) model $I^o$ in a
scale-less basis (delta functions). Such algorithms also require
regularizes to avoid over-fitting (which results into spurious compact
sources in the image). Usually, this regularization is done via a user
defined maximum number of components or/and global estimate of the
noise in the image. Extended emission, which is at a very different
scale than a compact component, is broken up into delta functions and
later smoothed to suppress the high frequency errors made in such a
representation. However since delta functions are at a scale smaller
than even the resolution element, this results into the well known
breaking-up of extended emission problem. In this paper we describe
an algorithm which decomposes the sky image into parameterized
Adaptive Scale Pixel (Asp) model. The parameters of the Aspen are
determined using non-linear minimization. The algorithm is sensitive
to the local spatial scale as well as the local signal-to-noise ratio.
\section{Algorithm}
The functional form for the Asp used in this paper is a symmetric two
dimensional Gaussian. The algorithm searches for the locally best fit
Asp to the Dirty Image, by estimating the location ($x_k,y_k$),
amplitude ($A_k$) and the size ($\sigma_k$) of the Asp.
The dirty image is smoothed to a few scales ranging from the smallest
to the largest expected scale. A global set of Aspen, $\{P_o\}$ is
maintained and a new Asp added to this list at each iteration. The
model image is computed using this set and Eq.~\ref{O2-3:MODIMG}. The
image decomposition into Aspen basis then proceeds as follows:
\begin{enumerate}
\item \label{ONE} At each iteration, compute the model and residual
visibilities as $V^R_i=V^M_i-V^o$, where $V^o$ is the observed
visibilities. Compute the residual image $I^R={\mathcal{FT}}(V^R)$
and smoothed versions at a few scales.
\item Locate the peak among all the smoothed versions of $I^R$. Use the location,
amplitude and the scale at which the peak was found as the initial
guess for the new Asp $P_k$. Set $\{P_o\}=\{P_o\}+P_k$.
\item Make a sub-set of Aspen $\{P_i\}$ which will
maximally affect the convergence.
\item Simultaneously solve for the parameters of this sub-set such that
$\chi^2=\sum |V^o-V^M_i|^2$ is minimized.
\item Goto Step~\ref{ONE} till $I^R$ is noise like.
\end{enumerate}
Ideally, all the Aspen determined in the earlier iterations should be
kept in the problem at each iteration. However since the
computational load scales strongly with the number of Aspen in the
problem, the speed of convergence is significantly improved by
adaptively dropping those Aspen which may not affect the $\chi^2$ at
the current iteration. Since the scale of the local Aspen is also
adaptively determined based on the local signal-to-noise ratio (SNR) and the
active set of Aspen determined in previous iterations are kept in the
problem, the number of effective parameters used to represent the
final image is minimized.
The set of active Aspen is determined by applying a threshold on the
length of the vector of the first derivatives of the $\chi^2$ with
respect to all the parameters ($p_i$) of each Aspen $P_k$
$D_k=\sum_i[(\partial\chi^2/\partial p_i)^2]^{1/2}$. $D_k$ is
computed for each $k$ at the start of each iteration and all Aspen with
$D_k