Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

504 J. Opt. Soc. Am. A / Vol. 22, No. 3 / March 2005 R. C. Flicker and F. J.

Rigaut

Anisoplanatic deconvolution of adaptive


optics images

Ralf C. Flicker and François J. Rigaut


Gemini Observatory, 670 North A’ohoku Place, Hilo, Hawaii 96720

Received April 7, 2004; accepted September 8, 2004


A modified method for maximum-likelihood deconvolution of astronomical adaptive optics images is presented.
By parametrizing the anisoplanatic character of the point-spread function (PSF), a simultaneous optimization
of the spatially variant PSF and the deconvolved image can be performed. In the ideal case of perfect infor-
mation, it is shown that the algorithm is able to perfectly cancel the adverse effects of anisoplanatism down to
the level of numerical precision. Exploring two different modes of deconvolution (using object bases of pixel
values or stellar field parameters), we then quantify the performance of the algorithm in the presence of Pois-
sonian noise for crowded and noncrowded stellar fields. © 2005 Optical Society of America
OCIS codes: 100.1830, 010.1080.

1. INTRODUCTION anisoplanatic PSF variation, and the second constrains


the characterization of the PSF variability to a low-
Use of adaptive optics (AO) on large astronomical tele-
dimensional space. The final requirement becomes nec-
scopes generally results in a field-varying point-spread
essary because, during the convergence of the algorithm,
function (PSF) in the compensated image. This property,
one must repeatedly update the PSF over the whole field.
known as anisoplanatism,1,2 leads to a gradual loss of
In addition, for the method to be of any practical use, we
sensitivity and resolution away from the AO system’s ref-
might add robustness and fast convergence as desired fea-
erence star. In addition, anisoplanatism also makes it
tures for the end product. The anisoplanatic convolution
very difficult to characterize the PSF over the whole field,
and the constrained PSF model are introduced in Section
which significantly compromises the accuracy with which
2. Section 3 outlines the maximum-likelihood deconvolu-
astrometric and photometric measurements can be made.
tion method and evaluate the derivatives that are re-
Although various image restoration schemes may be em- quired for use with a gradient-based search algorithm.
ployed to improve the image quality, most deconvolution In Section 4 some implementational details are discussed
methods and photometric techniques rely on the PSF be- and sample numerical results presented, and our conclu-
ing spatially invariant (isoplanatic). Previous ap- sions are summarized in Section 5.
proaches to deal with anisoplanatism in AO images dem-
onstrated some improvements by use of analytical models
and parametrizations.3–5 A novel method for 2. PRELIMINARIES
anisoplanatic deconvolution was recently studied by A. Convolution with a Space-Varying Kernel
Lauer,6 who incorporated a space-varying kernel by the Denoting image space and object space coordinates by x
method of Alard and Lupton7 into the standard ⫽ (x, y) and x⬘ ⫽ (x ⬘ , y ⬘ ), we write the image formation
Richardson–Lucy recursion.8,9 The method did not allow equation for an incoherent source as
for the PSF to be optimized, however; and as the algo-
rithm was applied directly to anisoplanatic AO data (the
galactic center with Hokupa’a and Gemini North), its per-
d 共 x兲 ⫽ 冕 dx⬘ h 共 x, x⬘ 兲 o 共 x⬘ 兲 ⫹ n 共 x兲 , (1)
formance could not be correctly quantified.
where d is the observed data, o is the object, h is the ker-
The myopic deconvolution method presented in this pa- nel of the object–image mapping (i.e., the PSF), and n is
per is essentially a fusion of the ideas discussed by Thié- an additional image noise. In the anisoplanatic case, the
baut and Conan,10 Lane,11 Magain et al.,12 Fusco et al.,13 kernel of this Fredholm integral equation of the second
and Lauer.6 Our contribution is chiefly the implementa- kind is not directly separable. Using a modal decompo-
tion of the anisoplanatic PSF optimization in a maximum- sition approach, however, we can obtain a partial separa-
likelihood framework. Constraints on the PSF result in tion by modeling it as a field-dependent sum of orthogonal
only a small number of free PSF parameters to be opti- functions6:
mized, hence the terminology myopic deconvolution (as
N⫺1
opposed to blind deconvolution, wherein the PSF is as-
sumed to be completely unknown). The following are h 共 x, x⬘ 兲 ⫽ 兺
i⫽0
a i 共 x⬘ 兲 p i 共 x ⫺ x⬘ 兲 , (2)
some of the basic requirements of the algorithm: (1)
space-varying kernel, (2) parametrized PSF variation, where N is the dimension of the discrete Hilbert space
and (3) efficient PSF computation. whereupon d and n are defined. The p i form an orthogo-
The first requirement is to accommodate the nal set of space-invariant PSF modes, and the a i are the

1084-7529/2005/030504-10$15.00 © 2005 Optical Society of America


R. C. Flicker and F. J. Rigaut Vol. 22, No. 3 / March 2005 / J. Opt. Soc. Am. A 505

associated amplitude fields that specify how the p i should izes the anisoplanatic convolution [Eq. (4)] and presup-
be weighted together over the image domain (see the ex- poses the existence of modes p i and amplitude fields a i for
ample in Fig. 1). Image formation can then be restated the first argument.
as Given M ⭐ N PSF estimations over the field of view,
one can form the matrix K whose columns are the indi-


N⫺1


vidual PSFs. A Karhunen–Loeve (KL) basis for this set
d 共 x兲 ⫽ dx⬘ a i 共 x⬘ 兲 o 共 x⬘ 兲 p i 共 x ⫺ x⬘ 兲 ⫹ n 共 x兲 (3)
i⫽0
may be obtained by diagonalizing the covariance matrix
C ⫽ K T K using, for example, singular value decomposi-
N⫺1 tion C ⫽ A⌳A T . This gives the orthogonal PSF modes
⫽ 兺
i⫽0
关共 a i o 兲 * p i 兴共 x兲 ⫹ n 共 x兲 , (4) p i as the columns of the matrix P ⫽ KA, and the ampli-
tude fields a i can be read directly off the rows of A. The
where the asterisk denotes ordinary convolution. For idea with the modal decomposition is twofold: (1) as the
brevity, we represent the mapping [Eq. (1)] by a bivariate amplitude fields a i vary smoothly over the field of view
operator using the symbol 䉺 so that d ⫽ h 䉺 o symbol- (cf. Fig. 1), the full-field amplitudes can be obtained by in-
terpolating a relatively sparse initial sample; and (2) the
KL basis will be dominated by a small subset of modes
with the highest singular values. Together, these two ob-
servations imply that we can obtain a sufficiently accu-
rate modal description by specifying only M Ⰶ N model
PSFs that cover the field adequately and then truncate
the expansion after m Ⰶ M modes at a conditioning num-
ber of 105 ⬃ 106 . This typically results in m ⫽ O(10),
whereas for small to large images N can range from
O(104 ) to O(106 ).

B. Parametrized Anisoplanatism Model


An ab initio analysis of how to model or reconstruct the
complete AO PSF from instrument statistical data is be-
yond the scope of this paper. A rather simplistic model is
used here, which forgoes most of the instrument-specific
details—as a model for the actual PSF of an instrument,
this would be far too crude; but for the purpose of evalu-
ating the performance of the deconvolution algorithm,
this level of realism will be sufficient. We will adopt the
von Karman model for the turbulence power spectral den-
sity (PSD) ⌽( ␬ ) ⫽ 0.023r 0⫺5/3( ␬ 2 ⫹ ␬ 02 ) ⫺11/6, where r 0 is
the Fried parameter, L 0 is the turbulence outer scale, and
␬ 0 ⫽ 1/L 0 with ␬ being the modulus of the spatial-
frequency vector ␬. We will further assume the layered
N l ⫺1
model C n2 (h) ⫽ 兺 l⫽0 c l ␦ (h ⫺ h l ) for the vertical turbu-
lence profile, where N l is the number of layers, the c l are
the strengths of the individual turbulence layers, and h l
are the layer altitudes. To first order, anisoplanatism
manifests only at the spatial frequencies that the AO sys-
tem attenuates. We model this partial compensation by
the spatial-frequency transfer function ⌸(␬). Ignoring
wave optical propagation effects, we can write the AO
anisoplanatism PSD as14
N l ⫺1

⌽ aniso共 ␬, ␪兲 ⫽ ⌸ 共 ␬兲 ⫻ 2⌽ 共 ␬ 兲 兺
l⫽0
f l 关 1 ⫺ cos共 2 ␲ h l ␪ – ␬兲兴 ,

(5)

where f l ⫽ 16.7c l ␭ ⫺2 r 05/3 are the layer relative turbulence


strengths (i.e., 兺 l f l ⫽ 1), ␪ is the angular distance from
the AO guide star, and ␭ is the wavelength. For the rest
of the AO system, we include only the isotropic PSD for
the fitting error, which we approximate by ⌽ fit ⫽ (1
⫺ ⌸)⌽. From the total PSD, one obtains the covariance
Fig. 1. First nine modes p i (top) and amplitude fields a i (bot-
by the Wiener–Kinchine theorem,15 and some further re-
tom) from a sample KL decomposition. The guide star is in the arrangement yields the structure function. Exponentiat-
lower left corner of the field (0, 0). ing produces the optical transfer function resulting from
506 J. Opt. Soc. Am. A / Vol. 22, No. 3 / March 2005 R. C. Flicker and F. J. Rigaut

anisoplanatism and AO fitting errors; and at this stage, needing to reiterate the cumbersome computational se-
transfer functions pertaining to other parts of the system quence to obtain new p i and a i each time. Adopting this
(detector, wind shake, static aberrations) are easily in- procedure, what is being fitted is in reality the isoplanatic
cluded when multiplied together. patch ␪ 0 of the model turbulence to the actual ␪ 0 of the
image. It is indeed seen in controlled simulations that,
C. Constrained Point-Spread Function Optimization under ideal conditions, the scale factor ␣ required to mini-
The anisoplanatism PSF model [Eq. (5)] introduced the mize the cost function (Section 4) is given by the ratio of
N l ⫺1
unknown parameters 兵 f l , h l 其 l⫽0 of the atmospheric tur- the two ␪ 0 ’s. The second constraint implies that the f l
bulence model, which must be estimated jointly with the are not optimized during the deconvolution. This may
image deconvolution (we assume that r 0 is known). sound like a severe restriction, but under typical observ-
There are a number of difficulties associated with this, ing conditions it is not. This is justified in part by the
however, that will require us to constrain the atmospheric fact that the observational sensitivity to many turbulence
model further. First, as evidenced by previous studies,16 parameters is in any case poor, and we are required to re-
several factors contribute to limiting the number of free duce the number of degrees of freedom to what may be ob-
parameters that can be observed and uniquely deter- served with statistical significance. Hence a generic ini-
mined from the anisoplanatism in astronomical AO im- tial distribution f l (or specific, as typical to the site) may
ages. In practice, we may therefore limit ourselves to be selected and kept fixed. Second, since ␪ 0 is a degen-
one- or two-parameter models. Second, the pixelized de- erate quantity that can be reproduced by various arrange-
convolution of d, which imposes no constraint on the ob- ments of f l and h l , there is no loss of generality when we
ject, cannot in general resolve the degeneracy between a adjust only h l to obtain the best-fit model ␪ 0 .
point source elongated by anisoplanatism and an ex- One limitation with this scheme of course is that it is
tended object with a lesser amount of anisoplanatism. not possible to extrapolate outside the initial field of the
Even though additional image constraints such as positiv- a i , so one must be sure to supply a sufficiently high ini-
ity, a good sky background subtraction, and prudent ini- tial guess that the algorithm will only need to interpolate.
tial conditions will lessen the effect of this degeneracy, we Hence, given a fixed distribution f l and an initial guess
are still prompted to consider alternatives to the pixelized for h l , the PSF optimization is reduced to a one-
deconvolution. One useful parametrization of the object parameter problem, that of determining the optimal scal-
o is the stellar field model, ing factor ␣. We introduce this dependency simply as
a i (x) → a i (x, ␣ ) ⫽ a i ( ␣ x).
N s ⫺1

o 共 x兲 ⫽ g 共 x兲 ⫹ 兺
k⫽0
␾ k ␦ 共 x ⫺ xk 兲 , (6)
D. Correct Sampling
Finally, we note that the deconvolution algorithm should
where xk and ␾ k are the stars’ positions and integrated not attempt to recover spatial frequencies beyond the cut-
photon counts, N s is the number of stars, and g(x) is a off frequency of the telescope or the instrument. Decon-
smooth function representing low-frequency background volving d with the total PSF h will result in a deconvolved
variations. Adopting the low-order polynomial expan- image that violates the Shannon sampling condition,15
sion g(x) ⫽ 兺 i, j c q x i y j , where q(i, j) is a meta-index, the which can lead to spurious artifacts being introduced into
set of stellar field parameters to be optimized jointly with the deconvolved image (by the so-called Gibbs phenom-
兵 f l , h l 其 are 兵 ␾ k , xk , c q 其 . This parametrization lifts the enon). Therefore one may use a narrower kernel s, cho-
pixelized degeneracy, permitting a unique solution. Fi- sen such that the deconvolved image ô has a residual PSF
nally, one computational complication in adjusting the r that does not violate the sampling condition.12 A
turbulence model on the fly is that, in principle, for every straightforward implementation of this idea that does not
change in a parameter (h l or f l ), one would have to go require the computation of s is to simply preconvolve d
through the whole process of PSF generation, covariance with r and then deconvolve the transformed problem d ⬘
matrix multiplication, and singular-value decomposition ⫽ h 䉺 o ⬘ , where d ⬘ ⫽ r * d and o ⬘ ⫽ r * o. For this pa-
to obtain new modes p i and amplitude fields a i . This per we used a Gaussian function for the residual PSF r,
would slow down the algorithm unacceptably, but it may with a full with at half-maximum (FWHM) close to that of
be circumvented with a few more constraints imposed on the diffraction-limited PSF. This low-pass filtering is in
the turbulence model: (1) initial h l fixed, and h l⬘ accord with the main concern of the deconvolution
⫽ ␣ h l ᭙l, ␣ 苸 R1 and (2) fixed distribution f l . method, which is merely to improve photometry by negat-
The first condition means that the layers are not al- ing anisoplanatic atmospheric effects and not attempting
lowed to move independently, but are constrained to be to retrieve information beyond the spatial cutoff fre-
comoving so as to preserve the morphology of the turbu- quency of the optical system.
lence profile. The scalar factor ␣ thus effectuates a lin-
ear stretch or compression of the profile. Together with
the second constraint (discussed further below), the angu-
lar turbulence correlation (i.e., the anisoplanatism) at a
given angle ␪ is preserved when h l and ␪ are scaled 3. MAXIMUM-LIKELIHOOD
jointly, so that statistically ( ␣ h l , ␪ / ␣ ) experiences the DECONVOLUTION
same anisoplanatism as (h l , ␪ ). This means that we Addressing deconvolution by way of function minimiza-
can follow a vertical scaling of the layer altitudes simply tion, we may state the problem as that of finding the op-
by scaling the modal amplitude fields a i laterally, without timal estimate o
*
R. C. Flicker and F. J. Rigaut Vol. 22, No. 3 / March 2005 / J. Opt. Soc. Am. A 507

o ⫽ arg min M共 ô 兲 (7) m⫺1


* ˆ o ⵜ k M ⫽ 2 ␣␾ k 兺i⫽0
ⵜa i 共 ␣ xk 兲共 ⑀ˆ * p i 兲共 xk 兲 (13)
that minimizes the cost function M, where ô is an esti-
m⫺1
mate of o. In maximum-likelihood deconvolution, the
cost function is commonly chosen as M(ô) ⫽ ⫺ln P(d兩ô), ⫹ 2␾k 兺 a i 共 ␣ xk 兲共 ⑀ˆ * ⵜp i 兲共 xk 兲 . (14)
where P(d 兩 ô) is the conditional probability of observing d i⫽0

given the estimate ô. We limit the scope here to the case We can avoid computing the gradients of p i by evaluating
of noise with Poissonian statistics, which can be either the convolutions in the Fourier domain and invoking ap-
image dependent or image independent. The former rep- propriate Fourier-transform identities. That is, with FT
resents photon noise, and the latter may be used to ap- denoting Fourier transform and the angular frequency
proximate detector noise, which is usually taken as vector u being the Fourier conjugate of x, we may use
Gaussian (for a given Gaussian width, a Possonian mean
can be found for which the distribution functions have FT共 ⑀ˆ * ⵜp i 兲 ⫽ 2 ␲ u冑⫺1 ⫻ FT共 ⑀ˆ 兲 ⫻ FT共 p i 兲 .
roughly the same shape).
The case of Gaussian noise is discussed briefly at the For the background function g(x) ⫽ 兺 q c q x i y j , the modal
end of this section, and a treatment was previously given derivative is
by Lane.11 For image-dependent Poisson noise, the like-
lihood is given by
N⫺1
d̂ 共 xi 兲 d 共 xi 兲
exp关 ⫺d̂ 共 xi 兲兴
⳵M
⳵cq
⫽ 冕 dx⑀ˆ 共 x兲共 h 䉺 x i y j 兲共 x兲 . (15)

P 共 d 兩 ô 兲 ⫽ 兿
i⫽0 d 共 xi 兲 !
, (8) As alluded to in Appendix A, this could have been ob-
tained directly from the Gateaux variation by setting v
where d̂ ⫽ h 䉺 ô is an estimate of d ⫺ n. The cost func- ⫽ x i y j . Upon including an image-independent Poisso-
tion evaluates to nian component of mean ␨, the above calculated deriva-
tives retain their form, so we need only modify d̂ → d̂
N⫺1
⫹ ␨ in the final expressions. It is possible to let the al-
M共 ô 兲 ⫽ ⫺ln P 共 d 兩 ô 兲 ⫽ 兺
i⫽0
关 d̂ 共 xi 兲 ⫺ d 共 xi 兲 ln d̂ 共 xi 兲兴 , gorithm optimize the value of ␨; however, this case was
covered already by the zeroth-order term of g(x), i.e., the
(9)
parameter c 0 for i ⫽ j ⫽ 0. The complexities of the
where terms not dependent on ô were omitted from the above computations are seen to scale linearly with the
sum. To minimize M with a gradient-based search algo- number of KL modes m and inherit the scaling of the fast
rithm, its functional and partial derivatives are required. Fourier transform with the image logical dimension.
For the pixelized case, the functional derivative of M They also scale linearly with the number of stars N s in
with respect to ô (see the derivation in Appendix A) is the parametrized stellar field model.
For completeness, we note here how to obtain the cor-
m⫺1
responding formulas for the case of image-independent
␦ M共 ô 兲 ⫽ 兺
i⫽0
a i 共 ␣ x兲共 ⑀ˆ * p i 兲共 x兲 , (10) Gaussian noise. With the standard deviation of the noise
being ␴, the Gaussian cost function evaluates to
where ⑀ˆ ⫽ (d̂ ⫺ d)/d̂. To enforce positivity of ô, we may N⫺1
apply the reparameterization ô ⫽ ␺ˆ 2 . 10 It is then 1
straightforward to show that ␦ M( ␺ˆ ) ⫽ 2 ␺ˆ ⫻ ␦ M(ô).
M共 ô 兲 ⫽
2␴ 2

i⫽0
关 d̂ 共 xi 兲 ⫺ d 共 xi 兲兴 2 , (16)

For the constrained PSF optimization, one must also com-


pute the partial ␣ derivative: where a constant term was omitted. The derivatives of
this summed-square error cost function are straightfor-


m⫺1
⳵M ward to evaluate, but they may also be conjured from the
⳵␣
⫽ dxô 共 x兲 兺
i⫽0
关 x • ⵜa i 共 ␣ x兲兴共 ⑀ˆ * p i 兲共 x兲 , (11) previous Poissonian case by simply redefining ⑀ˆ ⫽ (d̂
⫺ d)/ ␴ 2 in the final expressions of Eqs. (10)–(15). The
where ⵜ ⫽ ( ⳵ x , ⳵ y ) is the two-element gradient operator Gaussian cost function and its derivatives can then be
and the dot means the vectorial product. For the case of used with good results when the stars’ brightnesses do not
the parameterized stellar field model [Eq. (6)], the partial vary by several orders of magnitude. With large varia-
derivatives with respect to ␾ k and ␣ are found to be ex- tions, however, the learning rates of the search algorithm
actly the same as Eqs. (10) and (11); however, they are must be adjusted for the brightest stars to ensure stabil-
evaluated only at the current star coordinates xk . That ity, which implies very slow convergence for the fainter
is, for ␣ we would instead compute stars. It is of course possible to give the stars individual
learning rates by some rather ad hoc normalizations;
N s ⫺1 m⫺1
⳵M however, the reciprocal of d̂ that is present in the deriva-
⳵␣
⫽ 兺
k⫽0
␾k 兺
i⫽0
关 xk • ⵜa i 共 ␣ xk 兲兴共 ⑀ˆ * p i 兲共 xk 兲 . tives of the Poissonian cost function automatically pro-
vides a normalization to this effect. For a large part of
(12)
trial cases, we observe much better stability properties
The gradient with respect to the stars’ positions is com- with the Poissonian cost function, and hence we will not
puted as be using the Gaussian cost function in this paper.
508 J. Opt. Soc. Am. A / Vol. 22, No. 3 / March 2005 R. C. Flicker and F. J. Rigaut

4. IMPLEMENTATION AND RESULTS 3. Source Identification


An issue for the parametrized stellar field mode of decon-
A. Algorithm Implementation
volution is the identification of point sources to include in
Implicit algorithms are most commonly employed to
the set. Given that its initial conditions will contain a
search for an optimal configuration of variables, i.e.,
number of false detections while missing some of the real
兵 ␺ (x i , y i ), ␣ 其 in the pixel-based mode or 兵 ␾ k , xk , c q , ␣ 其
objects, the algorithm must be able to prune false detec-
in the parametrized stellar field model. The conjugate
tions and discover new sources as convergence progresses.
gradient (CG) algorithm is generally considered to be one
Rather than trying to compute the Hessian or explicit sa-
of the most efficient and reliable methods for function
liences, we employ simple thresholding criteria for the
minimization when the partial derivatives are known.17
pruning; typically a minimum photon count of a few sig-
The lack of a common Hilbert space for the ␣ and ␺ (or ␾ k )
mas above the background level and a nearest-neighbor
variables, however, tends to render a joint CG minimiza-
threshold of 0.5–1 pixels. Including new stars in the set
tion unstable. Although running separate but simulta-
can be done either in batch after initial convergence or se-
neous CG minimizations on ␣ and ␺ does improve stabil-
quentially during convergence by targeting the peak dis-
ity, this becomes instead detrimental to convergence by
the conjugate property of the ␺ search directions. Essen- crepancies of the residual image d ⫺ d̂. To obtain initial
tially, if the initial guess for ␣ is far from its best-fit value, conditions for as complete a set of stars as possible, we
search directions for ␺ expended during the early conver- suggest running the pixelized deconvolution as a prepro-
gence phase will introduce image errors that can be costly cessing step for the parametrized stellar field deconvolu-
to repair since the previous search directions may not be tion.
reused. Our observation is that these complications fre-
quently cause the CG to terminate (zero convergence B. Sample Numerical Results
rate) far from the optimal solution, or converge at a To evaluate the performance of the algorithm under vari-
grossly suboptimal rate. For these reasons we opted to ous conditions, we used a square grid of 16 ⫻ 16 (M
work chiefly with a steepest-descent search algorithm for ⫽ 256) model PSFs, retaining the first m ⫽ 12 modes of
this optimization problem. Other methods, such as the KL expansion throughout. First we investigate the
Levenberg–Marquardt and quasi-Newton methods17 have validity of the algorithm and the accuracy of the PSF op-
not yet been tested extensively, though there are indica- timization under ideal conditions. We omit noise and
tions that these might suffer from the same stability prob- avoid stellar crowding by constructing an artificial grid of
lems as CG. The simplicity of steepest-descent algorithm well-separated stars within a narrow magnitude range,
enables an explicit handle on the feedback, where we may 16 ⭐ M k ⭐ 18, where M k ⫽ ⫺2.5 log ␾k ⫹ Z is the
introduce various levels of control (e.g., momentum terms, K-band (2.2-␮m) magnitude and Z is the photometric zero
temporal ramping, sigmoid squashing functions, adaptive point. Conditions are then rigged so that anisoplanatism
learning rates). causes the Strehl ratio to vary from 0.75 at the guide star
(lower left corner, see Fig. 1) to 0.16 at the opposite corner
of the field. The convergence of the parametrized stellar
field deconvolution in the ideal case (lightface curves) and
1. Edge Wrapping with PSF fitting errors included (bold curves) are shown
If the fast Fourier transform is used to compute the con- in Fig. 2. The various quantities plotted are the cost
volutions, the imposed periodicity may cause flux from function M, (solid curves), the image error metric E
stars close to an edge of the image to be wrapped around (dashed curves)


to the opposite edge. This is avoided by defining an ex-
tended support for ô and padding d with the estimated
dx兩 d ⬘ 共 x兲 ⫺ d̂ ⬘ 共 x兲 兩 2
background level. If the background is field varying, an
E⫽

iterative approach may be employed: First deconvolve d , (17)
unpadded with respect to g(x) to obtain a first estimate dx兩 d ⬘ 共 x兲 兩 2

for the c q coefficients; then extend the support of d with


the estimated g(x) and deconvolve again. the anisoplanatism fitting error 兩 ␣ ⫺ ␣ˆ 兩 , (long dashed
*
curves), and the field-averaged photon count E ␾ (dotted
curves) and position E x (dotted–dashed curves) errors
N s ⫺1 N s ⫺1
2. Saturation 1 1
In deep AO images, it is usually the case that the guide E␾ ⫽
Ns

k⫽0
ˆ 兩,
兩␾k ⫺ ␾ k Ex ⫽
Ns

k⫽0
兩 xk ⫺ x̂k 兩 .
star is heavily saturated, and hence also any other bright
sources within the field. One option to deal with this is (18)
to simply mask out the saturated stars with a binary im- When the same turbulence model and KL expansion is
age plane mask. This is likely to leave a bright halo un- used for the model PSFs as for the image PSFs (lightface
accounted for, however, that will contaminate the decon- curves), the performance of the algorithm is observed to
volution of any faint stars within the envelope of the halo. be unbiased and its accuracy limited by numerical preci-
A better method is to introduce a clipping of ô at the same sion. The second set of curves in bold show the corre-
level as in d; but this needs to be accompanied by rela- sponding results upon inclusion of the fitting of a three-
tively good magnitude estimates for the clipped stars or layer model turbulence profile to a five-layer true
the bulk of the convergence time will be spent working atmosphere and a truncation of the KL series for the PSF
down these errors. model.
R. C. Flicker and F. J. Rigaut Vol. 22, No. 3 / March 2005 / J. Opt. Soc. Am. A 509

the dots (i.e., the individual stars) are the median and
one-sigma deviations as computed from histograms of the
data (solid curves with pluses) and the one-sigma devia-
tions for a theoretical model of ideal aperture photometry
(solid curve). For the aperture photometry model, we
used ⌬m ⫽ ⫺1.08/SNR (signal-to-noise ratio), with SNR
⫽ S ␾ /(S ␾ ⫹ b) 1/2, where b is the background photon
count in the aperture and S is the Strehl ratio. A single
common aperture size (11 pixels), which is nonoptimal but
representative, was chosen for the comparison, and the

Fig. 2. Convergence of the parametrized stellar field deconvolu-


tion in the ideal case (lightface) and with PSF fitting errors in-
cluded (bold). The quantities plotted are the cost function M
(solid curves), the normalized image error E (dashed curves), the
␣ error (long dashed curves), the field-averaged photon count E ␾
(dotted curves), and position E x (dotted–dashed curves) errors
reported in photons and pixels on the common ordinate.

Fig. 3. Simulation of 200 images with independent realizations


of photon noise and star magnitudes, deconvolved with the pa-
rametrized stellar field model. Curves indicate the median fit
and the one-sigma deviations as computed for the data (solid
curves with pluses) and as predicted by a theoretical model (solid
curves) for a Strehl ratio of 0.2.

The next set of simulations presented in Figs. 3–6 have


a similar degree of anisoplanatism (0.18 ⬍ Strehl
⬍ 0.75), but also include image-dependent photon noise
and a sky background level of 13.5 magnitudes/(arcsec)2
at K. We then emulate 200 separate 20-min K-band ex-
posures with an 8-m telescope (random noise and star
magnitude realizations), where the stars have K-band
magnitudes in the range 14 ⭐ M k ⭐ 22 (still no crowd-
ing). The stopping criteria for the algorithm was based Fig. 4. Data from Fig. 3 seperated into domains of the Strehl
ratio. Curves indicate the median fit and the one-sigma devia-
on the convergence rate of the cost function M, which we tions as computed for the data (solid curve with pluses) and as
calculated as the normalized standard deviation of M for predicted by a theoretical model (solid curve) for the mean Strehl
three consecutive iterations. The curves plotted on top of ratio of each bin. Note the varying scale of the ordinates.
510 J. Opt. Soc. Am. A / Vol. 22, No. 3 / March 2005 R. C. Flicker and F. J. Rigaut

acceptable, the limiting magnitude of the method can be


increased in this way by correcting for eventual biases.

5. CONCLUSIONS
We have investigated a method for deconvolving AO im-
ages that are significantly affected by anisoplanatism.
With a modal PSF representation and a one-parameter
anisoplanatism model, an efficient myopic deconvolution
algorithm is obtained that simultaneously optimizes the

Fig. 5. Simulation of 200 images with independent realizations


of photon noise and star magnitudes, deconvolved with the pix-
elized method. Curves indicate the median fit and the one-
sigma deviations as computed for the data (solid curve with
pluses) and as predicted by a theoretical model (solid curve) for a
Strehl ratio of 0.2.

Strehl ratio chosen from the lower limit of the bins.


While both algorithms are afflicted with some nonlinear
effects toward the faint end, the parametrized algorithm
is seen to do significantly better than the pixelized
method over all. While the behavior with respect to the
Strehl ratio is reasonably consistent in the parametrized
case, it is clear that convergence effects are present in the
pixelized results.
The final set of deconvolutions presented in Figs. 7–9
were done on a moderately crowded simulated stellar
field, with 785 stars over 512 ⫻ 512 pixels (left half of
Fig. 7). The anisoplanatism is somewhat milder here,
varying from 0.6 (lower left corner) to 0.25 (top right)—in
all other respects the simulation parameters were kept
the same as is previous cases. A logarithmic stretch of
the pixelized deconvolution is shown in the right-hand
panel of Fig. 7, with sample PSF cross sections plotted in
Fig. 8. As evident in both Figs. 7 and 8, the amount of
residual halo consistently increases as the Strehl ratio de-
creases. Nevertheless, the concentration of flux into the
core and corresponding halo suppression still increases
the contrast ratio of most stars by more than two orders of
magnitude. Applying the parametrized algorithm to the
same data, we observe remaining photometric and astro-
metric errors as plotted in Fig. 9. Outliers are identified
as cases of extreme crowding, but the greater concern is
the limiting magnitude beyond which the errors become
too large and, more troubling still, biases become signifi-
cant. Although it would be possible from this particular
simulation to say that, for instance, the photometry is un-
biased brighter than K ⫽ 20 and has a 1-␴ uncertainty of
0.15 magnitudes at that limit, it is not a simple matter to
predict limiting magnitudes in the general case. Nonlin-
ear behavior toward the faint end may be investigated by
injecting artificial stars into the image and deconvolving. Fig. 6. Data from Fig. 5 seperated into domains of the Strehl
ratio. Curves indicate the median fit and the one-sigma devia-
If systematic deviations are observed for the test stars, tions as computed for the data (solid curve with pluses) and as
the deconvolved magnitudes of the real stars can be ad- predicted by a theoretical model (solid curve) for the mean Strehl
justed accordingly. If a greater zero-mean uncertainty is ratio of each bin. Note the varying scale of the ordinates.
R. C. Flicker and F. J. Rigaut Vol. 22, No. 3 / March 2005 / J. Opt. Soc. Am. A 511

Fig. 7. Pixelized deconvolution (right) of a simulated 512 ⫻ 512 K-band image (left). The field of view is 10 arcsec for the full image
(bottom) and 2.2 arcsec for the closeups of the top panel. The image stretch is logarithmic.

Fig. 8. Cross sections of the stars indicated in Fig. 7, where


pluses show the original profile and diamonds show the profile of
the deconvolved star. In the upper panel the cut is along the x
axis, and in the lower panel are shown the cut along the guide
star vector (e.g., the radial direction) (solid curves) and the or-
thogonal (azimuthal) direction (dashed curves).

image and the anisoplanatic PSF. In the ideal case of Fig. 9. Magnitude (upper panel) and position (lower panel) er-
perfect information, the maximum-likelihood algorithm is rors (in arc seconds) for the parametrized deconvolution of the
K-band image in Fig. 7.
seen to be limited only by numerical precision. Simula-
tions demonstrate how the performance of the algorithm
depends on both the signal level and the Strehl ratio in gence effects. Given the large number of parameters
the parametrized case and, in the pixelized case, conver- that bear on the final quality, a limiting magnitude given
512 J. Opt. Soc. Am. A / Vol. 22, No. 3 / March 2005 R. C. Flicker and F. J. Rigaut

an error tolerance cannot be generally stated. For stellar


冋 册
N⫺1
共 h 䉺 v 兲共 xi 兲
photometry the parametrized mode will always be supe-
rior to the pixelized, as it is much more forgiving to PSF
␦ M共 ô; v 兲 ⫽ 兺
i⫽0
共 h 䉺 v 兲共 xi 兲 ⫺ d 共 xi 兲
共 h 䉺 ô 兲共 xi 兲
errors and essentially cannot be overtrained. For good (A3)
results with the pixelized algorithm, it is critical to have a N⫺1


good background subtraction, and an iterative approach
to finding the proper ␣ value might be advisable (for in- ⫽ ⑀ˆ 共 xi 兲共 h 䉺 v 兲共 xi 兲 , (A4)
i⫽0
stance, if there are a few field stars, use these with the
parameterized algorithm to obtain an initial estimate for where we defined ⑀ˆ ⫽ (d̂ ⫺ d)/d̂ and d̂ ⫽ h 䉺 ô. This is
␣). a general result that can be used to compute a modal de-
A direct comparison to other deconvolution methods rivative for any mode v. To compute a pointwise func-
have not yet been undertaken. The main benefit of the tional derivative at y, we set v(x) ⫽ ␦ (x ⫺ y) and first
method presented here is the reduced complexity of com- evaluate the term h 䉺 v:
puting field-varying PSFs when anisoplanatic effects


must be taken into account, which renders the computa- m⫺1

tion efficient on large images. Future work include im-


provements in the treatment of saturation effects, a
h 䉺 ␦ 共 x ⫺ y兲 ⫽ 兺
i⫽0
dx⬘ a i 共 ␣ x⬘ 兲 ␦ 共 x⬘ ⫺ y兲 p i 共 x ⫺ x⬘ 兲

smarter algorithm for pruning and star detection, and in- m⫺1
vestigating search algorithms enhancements such as
step-size adaptation and reinforcement learning.
⫽ 兺
i⫽0
a i 共 ␣ y兲 p i 共 x ⫺ y兲 . (A5)

Substituting back into Eq. (A4) and converting the sum-


APPENDIX A: DERIVATION OF ␦M mation to an integration gives the final result:
To illustrate the techniques that were employed in the m⫺1
various differentiations in this paper, we outline the deri-
vation of ␦ M(ô) in this appendix. The first Gateaux ␦ M共 ô 兲 ⫽ 兺
i⫽0
a i 共 ␣ y兲共 ⑀ˆ * p i 兲共 y兲 . (A6)
variation of M at ô in the direction v is given by
By precomputing a i and the Fourier transforms of p i ,
M共 ô ⫹ ␭v 兲 ⫺ M共 ô 兲
␦ M共 ô; v 兲 ⫽ lim , (A1) this derivative can be efficiently computed.
␭→0 ␭

where v ⫽ v(x) is any admissible function. Using the


cost function given in Eq. (9), we obtain
ACKNOWLEDGMENTS

再兺 冎
This reseach is supported by the Gemini Observatory,
N⫺1
1 which is operated by the Association of Universities for
␦ M共 ô; v 兲 ⫽ lim 关 h 䉺 共 ô ⫹ ␭v 兲兴共 xi 兲 Research in Astronomy, Inc., on behalf of the interna-
␭→0 ␭ i⫽0
tional Gemini partnership of Argentina, Australia, Brazil,

冉兺 冊
N⫺1 Canada, Chile, the United Kingdom, and the United
1
⫺ lim d 共 xi 兲 ln兵 关 h 䉺 共 ô ⫹ ␭v 兲兴共 xi 兲 其 States.
␭→0 ␭ i⫽0

冉兺
N⫺1
1 REFERENCES
⫺ lim 兵 共 h 䉺 ô 兲共 xi 兲
␭→0 ␭ i⫽0 1. D. L. Fried, ‘‘Anisoplanatism in adaptive optics,’’ J. Opt.


Soc. Am. 72, 52–61 (1982).
2. R. J. Sasiela, ‘‘Strehl ratios with various types of
⫺ d 共 xi 兲 ln关共 h 䉺 ô 兲共 xi 兲兴 其 anisoplanatism,’’ J. Opt. Soc. Am. A 9, 1398–1406 (1992).
3. E. Diolati, O. Bendinelli, D. Bonaccini, L. M. Close, D. G.
Currie, and G. Parmeggiani, ‘‘Starfinder: an IDL GUI-
N⫺1 based code to analyze crowded fields with isoplanatic cor-
⫽ 兺
i⫽0
共 h 䉺 v 兲共 xi 兲 recting PSF fitting,’’ in Adaptive Optical Systems Technol-
ogy, P. L. Wizinowich, ed., Proc. SPIE 4007, 879–887
(2000).

再 冋 册冎
N⫺1 4. T. Fusco, J.-M. Conan, L. M. Mugnier, V. Michau, and G.
1 共 h 䉺 v 兲共 xi 兲
⫺ 兺
i⫽0
d 共 xi 兲 lim
␭→0 ␭
ln 1 ⫹ ␭
共 h 䉺 ô 兲共 xi 兲
. Rousset, ‘‘Characterization of adaptive optics point spread
function for anisoplanatic imaging. Application to stellar
field deconvolution,’’ Astron. Astrophys. Suppl. Ser. 142,
149–156 (2000).
The remaining limit is of the form a(␭)/b(␭) with both
5. T. Fusco, L. M. Mugnier, J. Conan, F. Marchis, G. Chauvin,
a(␭) → 0 and b(␭) → 0 as ␭ → 0. Invoking l’Hôpital’s G. Rousset, A. Lagrange, D. Mouillet, and F. J. Roddier,
rule, the limit may be evaluated according to ‘‘Deconvolution of astronomical images obtained from
ground-based telescopes with adaptive optics,’’ in Adaptive
a共 ␭ 兲 a ⬘共 ␭ 兲 共 h 䉺 v 兲共 xi 兲 Optical System Technologies II, P. L. Wizinowich and D.
lim ⫽ lim ⫽ . (A2) Bonaccini, eds., Proc. SPIE 4839, 1065–1075 (2003).
␭→0 b 共 ␭ 兲 ␭→0 b ⬘ 共 ␭ 兲 共 h 䉺 ô 兲共 xi 兲 6. T. Lauer, ‘‘Deconvolution with a spatially-variant PSF,’’ in
Astronomical Data Analysis II., J.-L. Starck and F. D.
The variation simplifies to Murtagh, eds., Proc. SPIE 4847, 167–173 (2002).
R. C. Flicker and F. J. Rigaut Vol. 22, No. 3 / March 2005 / J. Opt. Soc. Am. A 513

7. C. Alard and R. H. Lupton, ‘‘A method for optimal image stellar fields,’’ Astron. Astrophys. Suppl. Ser. 134, 193–200
subtraction,’’ Astrophys. J. 503, 325 (1998). (1999).
8. W. Richardson, ‘‘Bayesian-based iterative method of image 14. F. J. Rigaut, J. Veran, and O. Lai, ‘‘Analytical model for
restoration,’’ J. Opt. Soc. Am. 62, 55–59 (1972). Shack–Hartmann-based adaptive optics systems,’’ in Adap-
9. L. B. Lucy, ‘‘An iterative technique for rectification of ob- tive Optical System Technologies, D. Bonaccini and R. K.
served distributions,’’ Astrophys. J. 79, 745–754 (1974). Tyson, eds., Proc. SPIE 3353, 1038–1048 (1998).
10. E. Thiébaut and J.-M. Conan, ‘‘Strict a priori constraints 15. R. N. Bracewell, The Fourier Transform and its Applica-
for maximum-likelihood blind deconvolution,’’ J. Opt. Soc. tions, 3rd ed. (McGraw-Hill, New York, 2000).
Am. A 12, 485–492 (1995). 16. R. C. Flicker and F. J. Rigaut, ‘‘Hokupa’a anisoplanatism
11. R. G. Lane, ‘‘Methods for maximum-likelihood deconvolu- study and Mauna Kea turbulence characterization,’’ Publ.
tion,’’ J. Opt. Soc. Am. A 13, 1992–1998 (1996). Astron. Soc. Pac. 114, 1006–1015 (2002).
12. P. Magain, F. Courbin, and S. Sohy, ‘‘Deconvolution with 17. W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P.
correct sampling,’’ Astrophys. J. 494, 472–477 (1998). Flannery, Numerical Recipes in C: The Art of Scientific
13. T. Fusco, J.-P. Véran, J.-M. Conan, and L. M. Mugnier, Computing, 2nd ed. (Cambridge U. Press, Cambridge, UK,
‘‘Myopic deconvolution method for adaptive optics images of 1992).

You might also like