Robust Statistics for Image Deconvolution

Matthias A. Leea , Tamás Budavárib , Richard L. Whitec , Charles Gulianb

a Department of Computer Science, Johns Hopkins University, Baltimore, MD, 21218 USA.
b Department of Applied Math and Statistics, Johns Hopkins University, Baltimore, MD,
arXiv:1711.02793v1 [astro-ph.IM] 8 Nov 2017

21218 USA.
c Space Telescope Science Institute, Baltimore, MD, 21218 USA.

Abstract
We present a blind multiframe image-deconvolution method based on robust
statistics. The usual shortcomings of iterative optimization of the likelihood
function are alleviated by minimizing the M -scale of the residuals, which achieves
more uniform convergence across the image. We focus on the deconvolution of
astronomical images, which are among the most challenging due to their huge
dynamic ranges and the frequent presence of large noise-dominated regions in
the images. We show that high-quality image reconstruction is possible even in
super-resolution and without the use of traditional regularization terms. Using
a robust ρ-function is straightforward to implement in a streaming setting and,
hence our method is applicable to the large volumes of astronomy images. The
power of our method is demonstrated on observations from the Sloan Digital
Sky Survey (Stripe 82) and we briefly discuss the feasibility of a pipeline based
on Graphical Processing Units for the next generation of telescope surveys.

1. Introduction

In the new era of astronomy surveys, dedicated telescopes observe the sky
every night to strategically map the celestial sources. The next-generation sur-
veys are capable of such high speed that repeated observations become possible,
opening a new window for research to systematically study changes over time.
The key requirement for time-domain astronomy is the development of sophisti-
cated algorithms that can maximize the information we gain from the data. The
image processing approaches we present in this paper are motivated by these
astronomical challenges but are not specific to such exposures and are expected
to work for long-range photography regardless of the content of the images.
Algorithmically deblurring single telescope images has had a long history.
At first during the 1970s, when Richardson [23] and Lucy [19] independently
introduced a Poisson-based iterative deconvolution method, generally known as

Email addresses: [email protected] (Matthias A. Lee), [email protected]

(Tamás Budavári), [email protected] (Richard L. White), [email protected] (Charles Gulian)

Preprint submitted to Elsevier November 9, 2017

the Richardson-Lucy, and then during the early days of the Hubble Space Tele-
scope when researchers rushed to remove the blur induced by the misaligned
optics Burrows et al. [5], Cunningham & Anthony [8], Krist & Hasan [16], Lucy
& Hook [18], Nunez & Llacer [22]. Most of these techniques rely on a variation
of an iterative deconvolution using a known point spread function (PSF). The
main difficulty with such methods is preventing numerical degeneracies caused
by the presence of noise and poorly constrained parameters [12]. Even more
challenging is when the PSF is unknown. Some techniques, such as the iterative
non-negative matrix factorization (NMF) approach described in Fish et al. [9],
have shown limited success simultaneously solving for the PSF and the latent
image. The aforementioned approaches are limited by the total amount of infor-
mation contained in the input image and the PSF (if available). Ground-based
imaging, astronomy or long-range photography a like, suffers from varying dis-
tortions introduced by turbulence in Earth’s atmosphere. Multiple exposures,
however, provide an opportunity to break the degeneracy of the varying PSFs
and the constant background image. Blind multiframe deconvolution uses mul-
tiple images to simultaneously solve for the latent image and the individual
PSFs. This blind deconvolution method is based on Fish’s approach of itera-
tively solving for both the PSF and the latent image, the difference being that
new observations are regularly introduced in order to add more information into
the deconvolution, this is the basis of the approach described in Harmeling et al.
[10].
The current state-of-the-art techniques in astronomy for extracting and com-
bining repeated observations are relatively simple methods when compared to
other areas of image processing. This is in part due to the incredibly high dy-
namic range of the images, as well as the occurrences of large noise dominated
area which tend to handicap methods from other areas. The two most common
techniques used are lucky imaging [26] and the coadding of observed images
[18, 2]. Both of these use linear combinations of the input images, which en-
ables proper noise propagation, although the full co-variance matrix is rarely
produced.
Lucky imaging is the process of only choosing the observations with the
very best PSF and throwing away the rest, often over 90% of the original data
is discarded. This yields sharper images, but with a low signal-to-noise ratio. A
Coadd image is the combination of multiple images, produced by generating a
pixel-by-pixel mean or median image, therefore suppressing noise and outliers.
However, considering that the input images have different PSFs, one needs
to first convolve them to match the worst acceptable blur before combining
the pixel values. This results in a high signal-to-noise ratio but also in an
overall blurrier solution than the majority of input images. In practice, usually
a combination of both of these methods are applied; only the best images are
chosen and then co-added, meaning that a large amount of potentially useful
data is discarded in the process.
In a perfect world we would like to clean up every single observation without
throwing away any useful data, and restore each of them to pristine condition
with infinite resolution, but this is impossible. It is, however, possible to extract

2
Figure 1: Shown above is an example of Sloan’s Stripe 82 observations, displaying the prob-
lematic features found in these images. Top shows an observation with low signal-to-noise and
a large PSF, Bottom same section of the SDSS Coadd Annis et al. [2], showing improvement
in signal-to-noise and definition of sources over the plain observation.

and combine information across all observed frames and produce a few sharper
images of higher resolution, exposing sources and features previously hidden in
blur and noise. True reconstruction is computationally complex and therefore
slow and laborious. In order to keep up with the growing volume of data, we
need fast, statistically sound tools to explore and deblur these images in near
real time.
For the purpose of this paper we demonstrate the application of our method
on the Sloan Digital Sky Survey (SDSS) [30], which over a span of 7 years
imaged a large part of the southern hemisphere with roughly 80 fold coverage.
This area known as Stripe 82 [1] is an ideal testbed for demonstrating the power
of our new methodology. The overall quality of Stripe 82 images varies widely,
see Fig. (1). Some images are blurry, some are faint and noisy, but most are a
combination of those. Annis et al. [2] produced a state-of-the-art coadd of this
region, which we use as a reference in our experiments. Due to the extremely
high dynamic-range we encourage the reader to view the included images on a
computer screen, as print materials have difficulty reproducing the full range of
contrast.
In this paper, we present a novel approach for extracting information from
atmospherically distorted repeated observations in order to produce a deblurred,
low-noise, higher resolution image. In section 2, we discuss previous work as well
as the general approach. In section 3 we introduce our robust improvements. In
section 4 we discuss the application, results and performance of our method and

3
finally, in 5 with conclude with future work and a summary of our achievements.

2. Deconvolution as likelihood optimization

Before we explore removing the blur, we must formally understand how

the blur is induced into our observations. For the mathematical description
we consider one-dimensional images represented by (column) vectors but all
derivations and equations are similar and apply in higher dimensions as well.
In particular here we will focus on 2-D images.
Our model is relatively simple: Each observation yt at time t, is modeled
as the convolution of the underlying true image, x and the PSF ft , on top of
which there is measurement noise t ,

yt = ft ∗ x + t . (1)

Our goal is to fit this model to all observations in time {yt }. Going forward
we will make the differentiation between the underlying true image, x, and its
estimate, x̃, likewise we will distinguish between the observation, yt , and our
reconstruction, ỹt . Considering that a convolution is a linear operation, we
can represent the 1D convolution with f by a matrix F , such that F x = f ∗x,
this also holds true for the equivalent 2D convolution. Hereafter we use this
convention to representing matrices with capital letters and vectors with lower
case symbols.
The literature features a variety of methods for deblurring with a known
point-spread function (PSF). Especially common are methods which maximize
the Poisson likelihood, such as the Richardson-Lucy [19, 23, 24] deconvolution
or the more noise resilient damped-Richardson-Lucy [28]. There are also blind
methods, e.g., Fish et al. [9], which do not require a known PSF, but instead
solve for it as part of the iteration. This is a crucial feature as in most applica-
tions, PSFs are not inherently known.
While the noise in CCD observations follows a Poisson distribution, cf. the
number of electrons in CCDs proportional to the number of photons in each
pixel, its applicability is often limited because creating a complete model for an
image is not practical due to many known and unknown contributions (gradi-
ents from moon, thin clouds, etc.) The image processing pipelines correct and
calibrate the input images. Estimates of the sky background are subtracted, and
while noise estimates for each pixel remain accessible, the transformed pixel val-
ues will have different noise properties. Fortunately the background counts in
typical images are high enough that a Gaussian likelihood is a good approxima-
tion to the likelihood function. Our approach originates in Bayesian statistics
but full inference is computationally too expensive and hence we resort to maxi-
mum likelihood estimation (MLE). The multi-frame blind deconvolution method
described by Harmeling et al. [10, 11] uses a quadratic cost function, which cor-
responds to the Gaussian limit. The resulting formula is similar to that of the
Richardson-Lucy. We can solve for the model image, x̃, which minimizes the

4
residuals in all pixels,
X 2
x̃ = arg min [y − F x] . (2)
x
pixels

In order to solve for the x̃-image update we use an adaptation of Harmeling’s

multiplicative update formula
x̃t+1 = x̃t  ut (3)
with
F T W yt
ut = (4)
F T W F x̃t
where  symbol and the fraction indicate element wise multiplication and di-
vision. We consider the fraction to be the update image, ut , derived from the
observation yt , which defines how adjustments are introduced into the model
image. This update image is then applied to our model, as shown in 3.
There are multiple ways to solve for the PSF, Harmeling et al. [10] proposes
using the LBFGS-B Algorithm [6], Homrighausen proposes a “Fourier deconvo-
lution” method [12] and Fish et al. [9] proposes simply using the same update
formula for the estimation of the PSF as for updating the model image. The
latter is possible since the convolution, f ∗x, is commutative, meaning that if
we can use an update formula to solve for the model, we must also be able
to rewrite it for the PSF. To do so, we simply interchange the occurrences of
the PSF and the model in our formula. Essentially we alternate holding the f
and the x̃ constant while estimating and updating the other. For the purposes
of this paper we assume the PSF to be constant across our images, but these
methods could be extended to use multiple separate PSF or use methods such
as Lauer’s approach to spatially-variant PSFs [17].
To regulate our solutions, we constrain both x̃ and f to be non-negative and
initialize them to non-zero values. While it is possible to initialize x̃ to a constant
value, we found initializing with the average of a few observations speeds up
convergence and yields better final results with fewer processed observations.
It is important to prevent erroneous zero values from cropping up in our
model and PSF, as once an element becomes truly zero, no future multiplicative
update will be able to change it. This is especially tricky as our intended
background, for our model image, is as close to zero as possible without exactly
reaching it. Floating point underflow can be part of this problem and therefore
needs to be dealt with appropriately.
In practice a number of problems occur due to random noise in the data and
numerical degeneracies. In the following sections we describe our novel methods
for overcoming these limitations.

3. Improving stability and convergence

3.1. Robust statistics
Arriving at a good estimate for the PSF image is crucial for successful decon-
volution. With the unmodified update formula, the cost function that is being

5
minimized is an L2 norm, therefore featuring a squared term. Extreme outliers
in the residual, especially early on in the estimation before the PSF has been
well formed, can overtake the residual, forcing convergence in those areas before
the rest of the image. This often leads to PSFs more resembling of a Dirac-δ
than a realistic PSF.
To curb the impact of those very large values, we borrow elements of robust
statistics and modify our quadratic cost function to be an M -estimator with a
robust ρ-function [20]. The solution is still iterative in nature and hence lends
itself well to our method. We adjust our cost function with the ρ-function,
X
x̃ = arg min ρ (y − F x) . (5)
x
pixels

In practice the solution to this general problem of robust statistics is obtained

by minimizing a weighted L2 problem
X 2
x̃ = arg min w [y − F x] , (6)
x
pixels

where the weight of each pixel w is computed for its residual based on the
current solution of plugging x̃ into W(r) = ρ0 (r)/r function [20],

w = W(y − F x̃) (7)

In particular we apply
 the bisquare function, also known as the Tukey bi-
weight, W(r) = min 3−3r2 +r4 , 1/r2 , which corresponds to a ρ -function that
is quadratic for low values but approaches a constant for large arguments, es-
sentially limiting the contribution of the largest residuals in the cost function.
Figure 2 illustrates the difference between the terms in our robust cost function.
For more details we refer the interested reader to the discussion of the bisquare
scale in [20].
The threshold at which the ρ-function deviates from being quadratic, is
tuned by scaling the residuals, which therefore controls where the dampening
starts. In order to relate this tuning parameter to the quality of the image, it
is natural to define this scaling in units of σ. This controlled dampening of the
residuals greatly increases the quality of the estimated PSFs and therefore also
the resulting image. More detailed results are presented in section 4.4.

3.2. Convergence
The addition of the weighting as described in the previous section helps con-
strain the PSF estimation and therefore the areas around objects. On the other
hand, the noise-dominated areas between objects remain under-constrained, re-
sulting in the occurrence of background artifacts.
The cause of these artifacts had been a long standing problem for this
method, until we noticed that as the deconvolution approaches convergence,
the useful and reasonable updates to our model image become smaller, more
sporadic and clustered around sources. Conversely, the updates in the regions

6
Figure 2: ρ-function associated with the bisquare family of functions. Note the dotted line
indicates the contribution of a purely quadratic function.

between sources, where values are already low and close to zero, can become ex-
treme due to noise present in the observed image. For example, if a background
pixel in our model is near zero, but noise in the observed image wants to push
the value higher, update factors of 1000x or more are not uncommon. In most
cases the resulting pixel value will still be very small and will fluctuate around
zero, cancel each other out. Unfortunately in some scenarios these persist as
artifacts, ie. bright speckles across the noise dominated areas of the image.
To prevent these updates from affecting the model image we introduce an
update clipping function,

u0t = max 1/d, min(d, ut )


(8)

where d > 1, which limits the maximum impact an update is allowed to have
on any single pixel. The closer the parameter d is to 1, the higher the impact
and therefore the more conservative the updates will be. When d is large, the
clipping has virtually no impact. This approach vastly cuts down the number
of background artifacts.
Limiting updates in such a way has no real drawback on the dynamic range
in practice. For d = 2 the contrast becomes 240 with only 20 iterations.
In fig. 3 we show an update image clipped with d = 2.0 (right) and the
corresponding current model estimate (left). Observe how the gray areas, where
the update is between 0.5 and 2.0, fall directly around the regions where objects
are located in the current estimate. The areas that either appear pure white or
black are where the update was originally above 2.0 or below 0.5 and therefore
have been clipped respectively, meaning we only allow each pixel to be either
doubled or halved with any update. The resulting effect of this clipping can be

7
 ;

Figure 3: A typical update (right) to our image model (left) contains the most reasonable
updates clustered around objects. The values between objects, the background, are already
near zero, so any noise or non-uniform background subtraction can produce extreme and
unreasonable update values. Note: the coloring of the update image is such that the areas
that fall below the bottom clipping are colored black and the areas that are above our clipping
range are colored white. The gray areas are values which we consider to be reasonable.

8
Figure 4: Updates to our model image fluctuate most in areas between sources, without
dampening these updates, speckles (left) get introduced into the noise-dominated background
and faint sources can get disrupted by extreme erroneous updates. By limiting the absolute
magnitude of these updates we get much more coherent result (right).

seen in fig. 4, the left image shows some typical artifacts, a speckled background
in noise dominated areas, as well as the adverse effects of extreme erroneous
updates to faint objects. On the right, is the result of an identical deconvolution
with the Update Clipping enabled.

4. Application
In the following sections we describe the more traditional components of our
method, as well as the overall algorithm. We present the results of our method
when applied to real data and also quickly discuss our technical implementation
and it’s performance.

4.1. Pixel censoring

Our linear model cannot represent omnipresent artifacts such as saturated
pixels, cosmic rays, etc. Masking these pixels allows us to process real images
containing these artifacts. We introduce binary mask images that zero out the
known-bad areas and leave the remainder untouched. Each incoming obser-
vation yt , has its own mask associated with it, which for our tests had been
pregenerated using masking data available from SDSS. Formally the mask ap-
plication can be written as a matrix multiplication with a diagonal matrix W ,
which enters the update formula as follows,
F T W yt
ut = . (9)
F T W F x̃t
Relatedly flux from objects located just beyond the edge of the image as
well as the abrupt edge of zero padding used with FFTs can cause artifacts

9
along the mask’s edges, which over multiple iterations can creep into the PSF
and begin corrupting the model. To prevent these artifacts, we taper the masks
towards zero around the edges and any masked out object, therefore smoothing
out these artificial hard edges.

4.2. Super Resolution

Access to multiple images of the same objects, also gives us the ability to
increase the spatial resolution beyond that of the source images. Based on the
nature of our images, we can expect each to have a slightly different relative sub-
pixel shift, meaning we can extract and combine that information into a higher
resolution image model. To do so, we introduce a downsampling operator B
that lowers the resolution of an image. If the model image x̃ has a factor of 2
higher resolution than the observations y, then B would halve the resolution of
x̃ in order to make it comparable to y for calculation of the residual. We can
model this relation as
y ≈ BF x̃ (10)
Similarly an upsampling operator can also be introduced, which is formally the
transpose, B T , such that BB T =I. These operators are easy to visualize when
B and B T act upon a vector. For a simple vector of length 3, the appropriate
operators would be as follows,
 
1  1 1
B=√  1 1
. (11)
2 
1 1

In practice, however, there is no need to create these matrices, we simply

implement an operation which doubles each pixel (in the 2x super resolution
case) along both dimensions and therefore grows the image. The inverse opera-
tion shrinks the image by a binning along each dimension, gathering up the flux
that was spread between pixels in the upsampling operation. Each upsampling
and downsampling operation is normalized such that the total sum of the fluxes
stays constant.
The addition of super resolution significantly increases the quality of our re-
sult; many small sources which in standard resolution resolve to small jagged/aliased
groups of pixels, become uniform in shape and better defined. Higher resolution
also allows us to more precisely measure sizes and distances between objects.

4.3. Assembling the pieces

The overall method proceeds as shown in Algorithm 1. For every observation,
yt , we initialize a new PSF and load any associated masks. We then iteratively
estimate a new PSF, while holding our image model constant. During this
iteration, we apply the aforementioned Robust Statistics weighting, repeatedly
updating the our PSF estimate until we’ve reached convergence. We use two
criterions to measure the PSF convergence, an absolute, which quantifies the
per-pixel change by calculating a root-mean-square of the differences between

10
the current and previous estimate, and a relative, which ensures we stop once the
maximum relative per-pixel change drops below 0.01%. We stop the convergence
once either of these criterions are violated. Once we have an acceptable solution,
we compute the update for the image model, while in turn holding the PSF
constant. The image model is only updated once per iteration, again we apply
our Robust Statistics weighting and our Update Clipping to prevent extreme
updates from having adverse effects on our model image. Since the resolved
PSFs can be wildly different between observations, the only piece of information
carried from one observation to the next is the image model.

Input : Repeated exposures, {yt }

Initialize latent image x̃ and the exposure masks;
for each new exposure yt do
Initialize PSF f˜t
while not converged do
Improve f˜t using robust statistics;
end
Solve for update ut with robust weights;
Apply clipping for stability of sky pixels;
Update estimated latent image x̃;
end
Output: Latent image and all PSFs
Algorithm 1: Robust blind deconvolution

4.4. Application to SDSS

As mentioned previously, we apply our method to repeated observations of
SDSS’s Stripe 82. We further restrict our study to the region where Stripe 82
overlaps with the Canada-France-Hawaii Telescope Legacy Survey (CFHTLS),
featuring a much larger telescope therefore yielding a much deeper image. This
allows us to make comparisons not only to the SDSS coadd but also the CFHTLS
coadd by Hudelot et al. [13]. CFHTLS has nearly double the angular resolution
of SDSS (SDSS pixel size 1px=0.396”, CFHT pixel size 1px=0.187”), making
it an ideal candidate for comparison of our super resolution results. CFHTLS
enables us to verify our results acting as a sort of ground truth. The particular
area we chose yields 68 usable observations. All comparisons figures in this
paper are shown on a log scale to highlight robust performance even close to
the sky background and around faint objects. For fairness of comparison we
scale all our images within each figure to the same total flux, contrast and bias.
In combination, the previously described methods allow us to produce supe-
rior results far exceeding the SDSS coadd in quality and even more impressively
the results of the CFHTLS coadd. In this section we will discuss our results
and the parameters used to achieve them. Good masking is the basis for all of
our results, we generate masks as described in section 4.1. The edges are then

11
Figure 5: unmodified multiframe blind deconvolution (MFBD) (top-left), MFBD with Update
Clipping (top-right), MFBD with robust statistics weighting (bottom-left), MFBD with robust
statistics weighting and update clipping (bottom-right). The clipping removes much of the
background noise, but does not have a large effect on the PSF. Robust Statistics weighting
provides a much improved PSF and the reduction of noise around objects.

softened using a 15 pixel Gaussian blur preventing edge artifacts. The next
ingredient is our robust statistics weighting, section 3.1. We experimented with
a large variety of different values for the tuning parameter and found T = 6 to
be a reliably good. The larger this parameter, the smaller the impact of the
robust statistics weighting, conversely as this parameter approaches 1, more and
more of the image will be affected and potentially causing bright sources to be
suppressed.
The sole application of the robust statistics weighting reduces the number of
artifacts and noise around and between objects, while also resolving a greatly
improved PSF, see bottom-left of fig. 5. Notice the lack of halos around the
objects, as well as the reduced number of the background speckles.
To further reduce these background artifacts and prevent erratic updates
from disrupting fainter sources, ie. the top center region of the image, we
introduce convergence control using the update clipping, section 3.2. In our
testing we found d = 2.0 to be a good clipping threshold. On its own (top-right
in fig. 5), the clipping clears up the vast majority of background speckles.
In combination these methods produce clean and sharp images, which far
exceed the quality of the SDSS coadd, which was generated using a similar set
of input image, and even exceeding the quality of the CFHTLS coadd. For a fair
comparison we show an excerpt of our results with and without super resolution
enabled to match the SDSS and CFHTLS coadds, see fig. 6.
Point sources are less challenging to deconvolve than complex sources, galax-
ies are a good benchmark for validating the quality of our PSFs, as there is more
structure that would otherwise get washed out by a poor PSF. In fig. 7 we show

12
Figure 6: Our standard (bottom-left) and super resolution (bottom-right) results show a clear
improvement in Signal-to-noise as well as a much smaller PSF as compared to both the SDSS
Coadd (top-left) and CFHTLS Coadd (top-right). Our super resolution result produces images
in comparable resolution and detail to CFHTLS, which is an impressive achievement given
that CFHTLS is a much deeper survey with more than double the resolution.

our robust performance on a spiral galaxy, note the additional detail available
on the spiral arms, also note the similarity in the structures of the CFHT coadd.

4.5. Software: Implementation and Performance

An important consideration with approaches like these is their performance.
If a great method takes many hours to process a set of small images, it may yield
great results, but it is impractical in a real world applications. For that reason
part of our focuses has been on implementing GPU-accelerated methods for all
of our compute intensive tasks. Originally we planned on implementing both a
full CPU and GPU code path for all features, but as we began to investigating
larger (2k by 2k) images, especially with super resolution, it became apparent
that a CPU implementation would be too slow to be useful.
Our implementation is written in Python, heavily relying on the use of
numpy [3], a fast numerical and array library, and pyCUDA [15], a Python
interface to NVIDIA’s GPU-programming language CUDA [21]. Python, while
perhaps not quite as performant as C or C++, offers access to an incredibly
rich and mature environment of existing scientific libraries, making development
and experimentation a pleasure. However, while investigating the addition of
wavelet filtering, we found a severe lack of functional GPU accelerated wavelet
transform libraries. At the time of writing there seems to be only one GPU-
accelerated python wavelet transform library, that being PyGASP [4], which
wildly under performs expectations, so much so that the well-known CPU-based
python wavelet library, pywt [27], outperformed PyGASP by 5x on the same
input. While we ultimately omitted wavelet filtering for this paper, we believe
it may be beneficial and is planned as future work.
Our current implementation can process 140 images (2k by 2k) using stan-
dard resolution in under 5 minutes and using 4k by 4k super resolution in ap-
proximately 10 minutes, testing performed on an NVIDIA K20 GPU. Of course
different parameter settings as well as varying quality of input image does affect

13
Figure 7: Deconvolution of complex sources such as galaxies are a good benchmark of how
well a PSF is formed. Here we compare our result (bottom-right) against a typical input frame
(top-left), as well as the SDSS (top-right) and CFHTLS (bottom-left) coadds.

14
the processing time. The times given here are based on our testing completed
with SDSS images.

5. Future Work and Summary

While our method performs well in many scenarios, we have identified a

variety of issues to be addressed as future work. In order for our method to
perform at it’s best, we rely on a uniform and known background level. Images
with poor background subtraction and/or non-uniform backgrounds can impact
the final result by wrongly adding or removing flux. At the moment these effects
can be largely mitigated by aggressively clipping our updates. Preferably we
would solve for a multi-variate background along side the PSF (as an additional
set of parameters), allowing us to appropriately remove the background. This
would further improve our results, by letting us dial back the clipping as well
as allowing us to include images that previously had to be discarded for large
background gradients. Another area of exploration is filtering and smoothing of
the residual in order to dampen noise. Initially we figured an approach based
on Starck and Murtagh’s wavelet filtering method [25] would work well, but
after the disappointing results with painfully slow wavelet transform libraries,
we decided to revisit this at a later time.
Additionally it would be beneficial to further validate our results against
other surveys such as the Dark Energy Survey (DES) [7] and the Hubble Legacy
Archive (HLA) [14, 29] verify the quality of our results. Also as part of our future
work we plan on evaluating our results in terms of changes to astronometric
and photometric errors. First results look promising, but more investigation is
needed.
Our method shows a vast gain in quality over Coadds, lucky imaging and
the standard Multiframe Blind Deconvolution. The robust statistics weighting
successfully prevents outliers to overtake the cost function, therefore providing
more realistic PSF estimates and much reduced background noise, the update
clipping is shown to prevent artifacts in noise dominated areas, as well as pre-
venting erratic updates from breaking up faint sources. In combination with
Super Resolution our methods outperform the current state-of-the-art for com-
bining images, allowing us to produce images that exceed even coadds of a much
larger telescope with twice the spacial resolution. We show how to produce ro-
bust results from real data in a timely manner. Our method will be instrumental
in tackling the colossal datasets produced by current and upcoming surveys, en-
abling the production of higher quality and depth of observations than initially
available from the instrument itself.

Acknowledgment

The authors acknowledge valuable initial discussions about blind deconvolu-

tion ideas with Michael Hirsch for discussions on related topics. This research
has been funded by NSF Grant AST-1412566.

15
References

[1] Abazajian, K. N., Adelman-McCarthy, J. K., Agüeros, M. A., Allam, S. S.,

Prieto, C. A., An, D., Anderson, K. S., Anderson, S. F., Annis, J., Bahcall,
N. A., et al. 2009, The Astrophysical Journal Supplement Series, 182, 543
[2] Annis, J., Soares-Santos, M., Strauss, M. A., Becker, A. C., Dodelson, S.,
Fan, X., Gunn, J. E., Hao, J., Ivezic, Z., Jester, S., et al. 2011, arXiv preprint
arXiv:1111.6619
[3] Ascher, D., Dubois, P. F., Hinsen, K., Hugunin, J., Oliphant, T., et al. 2001,
Numerical python
[4] Bowman, N., Carrier, E., & Wolffe, G. 2013, in Electro/Information Tech-
nology (EIT), 2013 IEEE International Conference on (IEEE), 1
[5] Burrows, C. J., Holtzman, J. A., Faber, S., Bely, P. Y., Hasan, H., Lynds,
C., & Schroeder, D. 1991, The Astrophysical Journal, 369, L21
[6] Byrd, R. H., Lu, P., Nocedal, J., & Zhu, C. 1995, SIAM Journal on Scientific
Computing, 16, 1190
[7] Collaboration, D. E. S., et al. 2005, arXiv preprint astro-ph/0510346
[8] Cunningham, C. C., & Anthony, D. 1993, Icarus, 102, 307
[9] Fish, D., Brinicombe, A., Pike, E., & Walker, J. 1995, JOSA A, 12, 58
[10] Harmeling, S., Hirsch, M., Sra, S., & Scholkopf, B. 2009, in Computational
Photography (ICCP), 2009 IEEE International Conference on (IEEE), 1
[11] Harmeling, S., Sra, S., Hirsch, M., & Scholkopf, B. 2010, in Image Process-
ing (ICIP), 2010 17th IEEE International Conference on (IEEE), 3313
[12] Homrighausen, D., Genovese, C., Connolly, A., Becker, A., & Owen, R.
2011, Publications of the Astronomical Society of the Pacific, 123, 1117
[13] Hudelot, P., Goranova, Y., Mellier, Y., McCracken, H. J., Magnard, F.,
Monnerville, M., Smah, G., Cuillandre, J.-C., et al. 2012, T0007: The final
cfhtls release
[14] Jenkner, H., Doxsey, R., Hanisch, R., Lubow, S., Miller III, W., & White,
R. 2006, in Astronomical Data Analysis Software and Systems XV, vol. 351,
406
[15] Klöckner, A., Pinto, N., Lee, Y., Catanzaro, B., Ivanov, P., & Fasih, A.
2012, Parallel Computing, 38, 157
[16] Krist, J., & Hasan, H. 1993, in Astronomical Data Analysis Software and
Systems II, vol. 52, 530
[17] Lauer, T. R. 2002, arXiv preprint astro-ph/0208247

16
[18] Lucy, L., & Hook, R. 1992, in Astronomical Data Analysis Software and
Systems I, vol. 25, 277
[19] Lucy, L. B. 1974, The Astronomical Journal, 79, 745
[20] Maronna, R., Martin, D., & Yohai, V. 2006, Robust statistics (John Wiley
& Sons, Chichester. ISBN)
[21] Nickolls, J., Buck, I., Garland, M., & Skadron, K. 2008, Queue, 6, 40
[22] Nunez, J., & Llacer, J. 1993, Publications of the Astronomical Society of
the Pacific, 1192

[23] Richardson, W. H. 1972, JOSA, 62, 55

[24] Starck, J., Pantin, E., & Murtagh, F. 2002, Publications of the Astronom-
ical Society of the Pacific, 114, 1051
[25] Starck, J.-L., & Murtagh, F. 1994, Astronomy and Astrophysics, 288, 342

[26] Tubbs, R. N. 2003, arXiv preprint astro-ph/0311481

[27] Wasilewski, F. 2010, Pywavelets: Discrete wavelet transform in python.
http://www.pybytes.com/pywavelets/
[28] White, R. L. 1994, in 1994 Symposium on Astronomical Telescopes & In-
strumentation for the 21st Century (International Society for Optics and Pho-
tonics), 1342
[29] Whitmore, B., Lindsay, K., & Stankiewicz, M. 2008, in Astronomical Data
Analysis Software and Systems XVII, vol. 394, 481
[30] York, D. G., Adelman, J., Anderson Jr, J. E., Anderson, S. F., Annis, J.,
Bahcall, N. A., Bakken, J., Barkhouser, R., Bastian, S., Berman, E., et al.
2000, The Astronomical Journal, 120, 1579

17