-
*[email protected]; phone 49 7243 992-252; fax 49 7243
992-299; www.fom.fgan.de
Blind Deconvolution Algorithms for the Restoration of
atmospherically
degraded Imagery: a comparative Analysis
Claudia S. Huebner*, Mario Greco
Dept. of Signatorics, FGAN-FOM, Gutleuthausstrasse 1, 76275
Ettlingen, Germany;
ABSTRACT
Suggestions from the field of image processing to compensate for
turbulence effects and restore degraded images include
motion-compensated image integration after which the image can
be considered as a non-distorted image that has been
blurred with a point spread function (PSF) the same size as the
pixel motions due to the turbulence. Since this PSF is
unknown, a blind deconvolution is still necessary to restore the
image. By utilising different blind deconvolution
algorithms along with the motion-compensated image integration,
several variants of this turbulence compensation
method are created. In this paper we discuss the differences of
the various blind deconvolution algorithms employed and
give a qualitative analysis of the turbulence compensation
variants by comparing their respective restoration results.
This
is done by visual inspection as well as by means of different
image quality metrics that analyse the high frequency
components.
Keywords: Image restoration, atmospheric turbulence, blind
deconvolution, image quality metrics
1. ITRODUCTIO
In imaging applications the prevalent effects of atmospheric
turbulence comprise image dancing as well as image
blurring. These image-degradation effects arise from random
inhomogeneities in the temperature distribution of the
atmosphere, producing small but significant fluctuations in the
index of refraction which are the most pronounced close
to the ground. Light waves propagating through the atmosphere
will sustain cumulative phase distortions as they pass
through these turbulence-induced fluctuations. When imaging over
horizontal paths, as opposed to vertical imaging, the
degree of image degradation is particularly severe. As a
consequence, image resolution is generally limited by
atmospheric turbulence rather than by design and quality of the
optical system being used.
A number of correction methods have been proposed over the
years, prominent among them the hardware-based
Adaptive Optics (AO) systems. Their underlying principle is to
measure the phase aberration of an incoming wavefront
and correct it, directly. Usually, only the higher Zernike modes
are corrected, like tip/tilt and defocus, by use of a
deformable mirror. Although they have proven invaluable for
imaging point sources in a variety of applications, e. g. in
astronomical or medical imaging, their effectiveness is somewhat
limited where extended targets are concerned. Efforts
have been made to remedy some of that drawback by using a hybrid
approach, i. e. by using soft-ware based methods in
addition to the hardware-based correction, e. g. in [1], either
directly "in-the-loop" or in post-processing. Nevertheless,
the work involved in the build-up of such a system is
comparatively complex [2], and the necessary equipment still
quite
expensive, not to mention bulky and immobile. On this account,
the relative simplicity of a software-based approach
which can be put into effect anytime and anywhere, at low cost,
holds great appeal for many applications, such as ours.
Our intention is, ultimately, the implementation of a mobile
turbulence-compensation system, which is able to correct
atmospheric image degradation effects in real-time on a
live-stream. Ideally, this correction should be put into effect
without necessitating the operator to undergo a protracted
initialization procedure while trying to find the optimal
setting
for a multitude of parameters. The work presented in this paper
could be considered as a step forward to that end.
1.1 Outline
Suggestions from the field of Image Processing to compensate for
turbulence effects and restore such degraded images include
Motion-Compensated Averaging (MCA), based on an idea of [3], after
which the image can be considered as a non-distorted image that has
been blurred with a PSF (Point Spread Function) the same size as
the pixel motions due to
Optics in Atmospheric Propagation and Adaptive Systems XI,
edited by Anton Kohnle, Karin Stein, John D. Gonglewski, Proc. of
SPIE Vol. 7108
71080M 2008 SPIE CCC code: 0277-786X/08/$18 doi:
10.1117/12.800124
Proc. of SPIE Vol. 7108 71080M-12008 SPIE Digital Library --
Subscriber Archive Copy
-
iii !CUI
the turbulence. Since this PSF is unknown, a (mostly) blind
deconvolution is still necessary to restore the image. We are
utilising four different algorithms in combination with this
motion-compensated image integration, namely the linear Inverse
Wiener Filter (IWF) method, the non-linear Lucy-Richardson
Deconvolution (LRD) along with its Iterative Blind Deconvolution
(IBD) variant which both use maximum likelihood estimation and,
last but not least, a deconvolution technique based on Principal
Component Analysis (PCA). The specifics of our implementation of
the motion compensated image integration are given in Section 2
along with some notions on the concept of image quality. In Section
3 the basic principle behind the blind deconvolution algorithms we
employed is briefly described and their respective advantages and
disadvantages are discussed. Our efforts to automate deconvolution
parameter estimation wherever possible, at least partially, are
elaborated on in Section 4 and in Section 5 the respective
restoration results of the turbulence compensation variants are
presented and evaluated.
(a) (b) (c)
Fig. 1. Downsized sample images from the sequence pairs with
different test patterns; for (a)-(c) left: 1.0 ms exposure
time,
right:
-
Search space
M()
by
1.3 Approach
It should be noted that during the development of the algorithm
presented in this paper, we did not adhere too rigidly to
the "real-time" objective in view of the continuing increase in
computational power available and with an additional code
optimisation in mind which will undoubtedly reduce computation
time yet a bit further. Nevertheless, we tried to keep
computation time per frame within reasonable limits, especially
considering that we are intending to process a stream of
images where the output rate should at least be a 30 fps (frames
per second) even if the input rate may comprise up to a
1000 fps. Naturally, such a constraint will come at the expense
of at least some of the quality of the results.
Consequently, we set out to find the best solution for our
application in terms of deciding which deconvolution method
yields the best results for a given choice of parameters within
reasonable time.
2. MOTIO-COMPESATED AVERAGIG
Motion Compensated Averaging (MCA) is in essence the same as
normal image integration, the main difference being,
that before integrating the next frame of the input sequence it
is shifted slightly within a given search space (see Fig. 2)
of a few pixels in every direction such that the input frame
better matches the running average. A sliding window of
length W is used for calculating this running average or,
alternatively, a temporal median as was proposed by [5]. The window
length ought to be increased with growing turbulence strength. The
maximum window length depends, of
course, on available memory space. Keeping 100 frames with a 240
256 resolution in the memory posed no problem,
as this was the setting we used, but keeping 1000 will (at
present) become problematic on most PCs.
Fig. 2. Illustration of the MCA algorithm
Increasing the search space, automatically also increases the
computational complexity by (2*+1) where denotes the maximum pixel
shift. Accordingly, the shift operation results in a number of
(2*+1) shifted images from which the one, that matches the previous
output image the best, needs to be determined. There are a number
of different approaches
to making that decision by employing quality metrics which will
be addressed in the following section. To keep calculation time
down, the size of the search space was limited to no more than two
or three pixels. Nevertheless,
calculation time per frame still is in the order of 0.1 sec for
a search space of 2 pixels and about 0.15 sec for a search
space of 3 pixels and therefore much too high to be real-time
capable. But given that hardware based implementations of
image shifting and stacking operations are comparatively easy to
realise, e. g. by using an FPGA (Field Programmable
Gate Array), the somewhat extended calculation time of our
software implementation shall be of no consequence here.
The idea of using MCA at all, apart from noise reduction which
would obviously also result from normal averaging, is
the reduction of image dancing between one frame and the next.
Unfortunately, in highly anisoplanatic conditions, such
as we are wont to face, the coherence length r0 of the
turbulence is much smaller than our aperture, so that the logic
doesnt quite apply. Nevertheless, when compared to the results from
normal averaging, the stabilization effect for the
results from the MCA is generally better.
2.1 IQM Image Quality Metrics
The problem of giving a qualitative measure about the degree of
similarity between two given images is closely related
to the concept of image quality and its many interpretations. As
such, there is no simple answer to the question about image
quality. For once, it strongly depends on the application at hand.
In image or video compression, for instance, it is
In Input
M(x) Shift
n (x) = In(x+M(x)) MC Input Sn+1 Current output
Sn Previous outputSn+1 = a Sn + (1 a) In Normal AverageSn+1 = b
[a Sn + (1 a) n ] + (1 b) In MC-Average
Proc. of SPIE Vol. 7108 71080M-3
-
comparatively easy to decide on an images quality. Here, a
higher compression rate is equivalent to lower image quality
and so-called full reference Image Quality Metrics (IQM) are
employed to compare the perfect, i.e. uncompressed, reference image
to its compressed instance. As our task is to decide upon the
closest match between two images, such an
IQM is the logical choice. But again calculation time is an
obvious issue and for this reason simple error summation
methods as the Mean Square Error (MSE), Root Mean Square Error
(RMSE) or Mean Absolute Error (MAE) are very
attractive. In [6] another universal IQM was suggested that
makes use of the statistical values of two images x and y measuring
structural distortion rather than the error energy. It is defined
by:
( ) ( ) 2y2xyx
yx
y x
221
2
yx
yx 2IQM
+= (2.1)
while x and y denote the respective mean values, x and y are the
standard deviations and xy is the cross-covariance. The basic idea
seemed promising, but so far we could detect no discernible
improvement in our application over simply
using the Absolute Error (AE), therefore we settled once more
for the fastest solution.
However, given that we are also interested in evaluating the
respective results of our restoration variants, there is
another
class of measures we need to consider, the so-called no
reference IQMs where no reference image is required. In the case of
evaluating restoration result such an ideal image obviously doesnt
exist. Considering, that our aim is to improve the
resolution of image data, good image quality will consequently
refer to high contrast of fine detail. Since the low-pass
filter effect of the atmosphere removes the high frequency
components, it is a logical conclusion to consider a metric
that
exploits this information. Some such metric was proposed by [7]
and has been employed by us before in the context of a
different project concerning a synthetic imaging technique. This
metric is the mathematical equivalent of an existing
physical system and is defined as:
= rr 242 d|)]}(IiFT{exp[|IQM (2.2) where FT{} denotes the
Fourier transform, I(r) the intensity values of the image I at
position r = [r1, r2].
3. BLID DECOVOLUTIO
Essentially, a deconvolution describes the procedure of
separating two convolved signals f and h. In the spatial domain the
blind deconvolution problem takes the general form:
( ) ( ) ( ) ( )yxnyxfyxhyxg ,,,, += (3.1) where g denotes the
blurred image, h the blurring or point spread function (PSF), f the
true image, * the convolution operator and n an additive noise
component. To simplify further steps, it is common practice to
transfer the problem into the Fourier domain where the relatively
complicated convolution-operation becomes a simple
multiplication:
( ) ( ) ( ) ( )vuvuvuvu ,N,F,H,G += (3.2) with G, H, F and N
denoting the Fourier transforms of g, h, f and n, respectively.
Many attempts at solving this deceptively simple equation for a
wide variety of applications can be found throughout literature. An
overview of the
most popular of these blind image deconvolution algorithms is
detailed in [8].
3.1 IWF Inverse Wiener Filtering
The simplest approach to a solution of Eq. (3.2) is to use
direct inverse filtering, i. e. to form an estimate F of the
undegraded image:
( ) ( )( ) ( ) ( )( )( )vu
vuvuvu
vu
vuvu
,H
,N,F,F
,H
,G,F +== (3.3)
It becomes obvious that blind deconvolution is, as such, an
ill-posed problem. Even if the exact blurring function was
known, due to the random noise component, the true image can
never be fully recovered. In addition, the filter function
H is likely to contain numerous zeros or at least near vanishing
values, such that the quotient on the right hand side of Eq. (3.3)
will produce significant errors in the restoration estimate, even
if noise as such was negligible. For increasing spatial frequencies
the ratio will also increase and thus disproportionately enhance
the relative effect of the noise.
Proc. of SPIE Vol. 7108 71080M-4
-
Deconvolution with an inverse Wiener filter uses the least mean
square error between the estimate and the true image:
( ){ }22 ffEerr = (3.4) where E{} denotes the expectation
operator. The Fourier transform of the estimate F can now be
expressed as [9]:
( ) ( )( ) ( ) ( ) ( ) ( )vuvuvuvuvuvu
vu ,G,F,N,H,H
,H,F
222
2
+= (3.5)
If the noise-to-signal power ratio ||/|F| becomes zero, meaning
the absence of additive noise, Eq. (3.5) will reduce to the ideal
inverse filter in Eq. (3.3).
3.2 LRD Lucy-Richardson Deconvolution
The Lucy-Richardson (LRD) algorithm was developed independently
by [10] and [11] and is a nonlinear and basically
non-blind method, meaning the PSF, or at least a very good
estimate, must be a priori known as is the case for the IWF.
The LRD has been derived from Bayesian probability theory where
image data are considered to be random quantities
that are assumed to have a certain likelihood of being produced
from a family of other possible random quantities. The
problem regarding the likelihood that the estimated true image,
after convolution with the PSF, is in fact identical with
the blurred input image, except for noise, is formulated as a
so-called likelihood function, which is iteratively maximized. The
solution of this maximization requires the convergence of [9]:
( ) ( ) ( ) ( )( ) ( )
=+
yxfyxh
yxgyxhyxfyxf kk
,,
,,,, 1 (3.6)
where k denotes the k-th iteration. It is the division by f that
constitutes the algorithm's nonlinear nature. The image estimate is
assumed to contain Poisson distributed noise which is appropriate
for photon noise in the data whereas
additive Gaussian noise, typical for sensor read-out, is
ignored. In order to reduce noise amplification, which is a
general
problem of maximum likelihood methods, it is common practice to
introduce a dampening threshold below which further
iterations are (locally) suppressed. Otherwise high iteration
numbers introduce artefacts to originally smooth image
regions.
3.3 IBD Iterative Blind Deconvolution
The iterative blind deconvolution (IBD) algorithm, proposed by
[12], is mainly a blind version of the LRD algorithm
where the PSF needs not to be known, only its support. The IBD
is a so-called Expectation Maximization (EM) algorithm which is an
optimization strategy for estimating random quantities corrupted by
noise. In the case of blind
deconvolution this means that the likelihood function from the
LRD algorithm is again maximized iteratively but with
specified constraints until an estimate for the blurring PSF is
retrieved from the data along with the estimate for the true
image. The IBD algorithm is characterized by a computational
complexity in the order of O( log2 ) per iteration where is the
total number of pixels in a single frame while normally more than
one iteration is required for its convergence.
3.4 PCA Principal Component Analysis
Due to the iterative nature of most blind deconvolution
algorithms, calculation time per frame is relatively high, thus
rendering any real-time applications virtually impossible. But
like the Inverse Wiener Filter, the deconvolution algorithm
based on Principal Component Analysis (PCA), which has been
proposed in literature only recently [13], is an explicit,
i.e. non-iterative, algorithm and executes very fast. The PCA
algorithm aims at de-correlating the correlation between
image matrix columns that has been caused by atmospheric (and
system) blurring. There exist two versions of it, a
"multiple-observation" version that operates on a sequence of
images, yielding a single output image, and functions as a
truly blind deconvolution algorithm and another
"single-observation" version that operates on single images and
needs
additional information about the filter support size.
If a number of M blurred observations of the same scene are
given they can be modelled as random vectors { }MYYY ,...,, 21 with
mean value Y . The PCA based algorithm looks for a linear
transformation or filter that maximizes the variance of its output,
i. e. its estimation for the true image, by exploiting the M
blurred input images as follows:
[ ] YYYY vYYYX += ,...,, 21 Mest (3.7)
Proc. of SPIE Vol. 7108 71080M-5
-
where estX is a random vector representing the estimated true
image obtained by this linear combination. The solution s to the
maximization problem is a vector of length M and can be written
as:
( ) ( ){ } maxarg YYv
XXs = estt
est (3.8)
with E{} denoting the expectation operator as before and v the
variance of the argument. This solution allows us to
extract the maximum spatial variance within the M-blurred input
images, thus recovering some of the high spatial frequency
components that have been removed previously by the blur.
The most obvious advantage of the PCA-based algorithm, apart
from being robust to white noise, is that according to
[13] its computational complexity is generally lower than that
of the iterative deconvolution techniques discussed in this
context.
3.5 Discussion
The four blind deconvolution algorithms discussed in this
section were chosen for a number of reasons. The IWF, for
instance, is a classic method which can be calculated directly
and therefore executes very fast, i. e. within 0.05 and 0.09
seconds for a filter size of 7 7 pixels (which was used for all
calculation time measurements). The results generally
improve if the process is repeated at least 2-3 times but with
the real-time application in mind that theoretically only
allows for a maximum of 0.03 sec for the complete processing of
a single frame, one calculation must suffice.
The IBD was mostly chosen for its general reliability and as a
kind of benchmark with which to compare the respective
results of the other methods. Since it usually takes several
iterations to converge, execution time for the 5 to 10
iterations
necessary to get a decent result is relatively long with 0.5 sec
up to 1.1 sec. The main reason to include the non-blind
LRD as well was as a means to evaluating the effectiveness of
our estimate for the point spread function in comparison
to the results yielded by passing only filter support size to
the IBD algorithm. Calculation time per frame for the LRD
was in the range of 0.36 sec and 0.45 sec for a number of 5 and
10 iterations, respectively, whereas 20 iterations already
took 0.97 sec.
The algorithm based on Principal Component Analysis was chosen
because it was claimed to be very fast and the
qualitative analysis in [14] which was done on synthetic data,
had given rise to hope that the algorithm would achieve
similarly good results when applied to real data. Calculation
time in the single-observation version proved to be about
0.18 sec per frame regardless of the number of principal
components to be determined. It took the multiple-observation
version 0.45 sec for 100 frames to calculate the first
component, containing the highest variance, 0.19 sec for 50
frames
and 0.03 sec for 10 frames. Calculation of all four principal
components took 0.62 sec for 100 frames, and for 50 and 10
frames calculation times were almost identical. Thus far, the
algorithm was not quite as fast as we had hoped for large
filter support sizes of 7 pixels and more.
It should be noted that all computations were done using Matlab
such that the given calculation times are relative and
might very well be sped up by an efficient
C/C++-Implementation.
4. PARAMETER ESTIMATIO
We tried automating the parameter estimation wherever possible
by reducing the input parameters for external conditions
to turbulence strength, characterized by Cn, the structure
parameter of the refractive index fluctuations, and approximate
path length. The input parameters for the optical system, like
frame-rate and image resolution, ought to be possibly be
communicated directly from the sensor without requiring manual
input as of now they are simply put down in a script
and read in.
4.1 Automatic oise Estimation
For the automatic noise estimation we decided upon using the
temporal variance of a pixel region with homogeneous
grey values. In order to find such a homogeneous region, we
first determined the image entropy of a single frame. Since
we were interested in local information content we performed the
calculations in blocks of 16 16 pixels. We then
picked the block with the lowest entropy and calculated the
temporal variance for the corresponding pixels over an
interval of a 100 frames (in pre-processing) and chose the mean
value of this variance as noise parameter. The resulting
values for the individual sequences have been listed in Table 1.
The intrinsic snag in this method is, of course, that the
variance caused by light which is randomly scattered due to the
turbulence fluctuations, becomes an inseparable part of
Proc. of SPIE Vol. 7108 71080M-6
-
illililt
S S a
:
1I
"It" "It" iiUiiLi1
I IILIMI41 I 11111R4
the result. It is also interesting to note, if perhaps not all
that surprising, that the values we derived this way differed
significantly from the one, namely 1.7519, we derived from a
sequence we recorded recently using the same camera
directed at a sheet of white paper, with the same image
resolution and exposure time but a different optics.
4.2 Estimation of the Point Spread Function
While the Optical Transfer Function (OTF) describes the response
of an optical system to a dirac impulse in the
frequency domain, its inverse Fourier transform, the so-called
Point Spread Function (PSF), characterises the degree to
which the optical system spreads, or rather, blurs the light
from a point source in the spatial domain. The approach to
image restoration involves the previous determination of the
atmospheric point spread function. The simplest
approximation to an atmospheric blurring function is a Gaussian
filter. This is quite a reasonable assumption since the
optical turbulence basically acts as a low-pass filter by
filtering out the high spatial frequencies and thereby blurring
sharp edges and point-like objects.
The main problem is to guess the correct filter size in terms of
mean and standard deviation : Choosing too big will introduce
ringing effects and choosing it too small will retain too much
blurring. For each of our sequences, statistics
regarding the turbulence conditions had been compiled and
evaluated with special focus on edge width broadening [15],
so we were able to use the edge width information that was
correlated with the Cn values that had been measured for our
respective sequences. The resulting estimates are listed in Table
1.
Seq.1, median 100 IBD median LRD median IWF median
Seq.2, median 100 IBD median LRD median IWF median
Seq.1, PCA single C 1234 for 100 PCA multi C 1234 for 100 Seq.2,
PCA single C 1234 for 100 PCA multi C 1234 for 100
Fig. 3: Some characteristic deconvolution results from Seq. 1
& 2 (7.15 a.m. & 7.20 a.m.), beginning with median of
100
frames as reference, followed by the resp. IBD, LRD and IWF
results (Seq. 1 top, Seq. 2 centre row) and concluding
with single and multiple version PCA results for components 1,
2, 3 and 4 (bottom row: Seq. 1 left, Seq. 2 right)
Proc. of SPIE Vol. 7108 71080M-7
-
AIVSI
AIVSI I
I
4AIMS
I I
5. RESULTS
For easier reference, the sequences will be numbered by time of
their recording, meaning Seq. 1 will refer to the earliest
recording from 7.15 a.m., Seq. 2 to the next, etc. And due to lack
of space all the images had to be reduced to 60% of their original
size. For Seq. 1-6 the respective average and median images of 100
frames look almost the same as the sample frame, merely a little
less noisy, so samples and averages were omitted and only the
median image, which was also input for the various deconvolution
algorithms, is displayed. Results are presented for sequence pairs,
so the image labels have been coloured to group results by
sequence. Fig. 3, 4 and 5 present a basic set of results for
sequences 1 & 2 (7.15 a.m. & 7.20 a.m., low turbulence), 3
& 4 (7.45 a.m. & 7.48 a.m., medium turbulence) and 5 &
6 (8.12 a.m. & 8.22 a.m., medium to strong turbulence),
respectively. This set includes the temporal median of 100 frames
(the same 100 frames in every case) as reference to compare to
restoration results, the respective IBD, LRD and IWF results,
followed by PCA single and multiple (only Seq. 1 & 2) version
results for various principal components, where "C 14", for
instance, means that components 1 and 4 have been evaluated by
exploiting the generalisation made in [14].
Seq.3, median 100 IBD median LRD median IWF median
Seq.4, median 100 IBD median LRD median IWF median
Seq.3, PCA single C 14 for 100 PCA single C 1234 for 100 Seq.4,
PCA single C 23 for 100 PCA single C 1234 for 100
Fig. 4: Some characteristic deconvolution results from Seq. 3
& 4 (7.45 a.m. & 7.48 a.m.), beginning with median of
100
frames as reference, followed by the resp. IBD, LRD and IWF
results (Seq. 3 top, Seq. 4 centre row) and concluding
with various single version PCA results (bottom row: Seq. 3
left, Seq. 4 right)
Due to lack of space only one comprehensive set of results could
be included. Exemplarily, the results for Seq. 7, i. e. the "worst
case scenario", were chosen, since they were the most significant.
They are presented in Fig. 6, which spreads over two pages,
beginning with a sample, median as well as average of 100 frames
and the total average of the sequence as frame of reference. To
illustrate the turbulence effect on the edges, the standard
deviation of the total sequence is also included. Subsequently, the
respective IBD, LRD and IWF results for both, average and median of
100 frames, are displayed; following next are results from the
single observation PCA deconvolution when applied to the average of
100, median of 20 and 10 frames and finally results from the multi
observation PCA applied to 10, 20 and 100 frames.
Proc. of SPIE Vol. 7108 71080M-8
-
-1rIII. _.iiill.ill ill
HI!.
Seq.5, median 100 IBD median LRD median IWF median
Seq.6, median 100 IBD median LRD median IWF median
Seq.5, PCA single C 1 for 100 PCA single C 1234 for 100 Seq.6,
PCA single C 2 for 100 PCA single C 1234 for 100
Fig. 5: Some characteristic deconvolution results from Seq. 5
& 6 (8.12 a.m. & 8.22 a.m.), beginning with median of
100
frames as reference, followed by the resp. IBD, LRD and IWF
results (Seq. 5 top, Seq. 6 centre row) and concluding
with various single version PCA results (bottom row: Seq. 5
left, Seq. 6 right)
5.1 Evaluation
5.1.1 Short exposure vs. long exposure
When imaging through turbulence, exposure time becomes an issue.
For longer exposures, i. e. if exposure time outlasts
the duration of the turbulence, turbulence-cells of all sizes
are averaged over. It is the largest cells which are
predominantly responsible for the image dancing as they move
across the aperture of the optics and the resulting
"smoke" effect which arises from light that gets deflected by
these larger cells. For short exposures, on the other hand,
these large cells are "frozen" and only the small ones, which
are responsible for the blurring, remain. Essentially, that
means that the geometry of a scene will be preserved, i. e.
straight lines will remain straight, but details will be
blurred.
Consequently, the results from the sequences with exposure below
0.1 ms were generally more accurate but also blurrier
than those with 1 ms exposure time which in the case of
sequences 2, 3, 5 and 7 definitely qualified for the long
exposure
case. The long exposure results looked sharper, had higher
contrast and contained more details.
5.1.2 Visual inspection
The deblurring efforts of IBD, LRD and IWF are barely
distinguishable and offer comparatively little improvement over
the median for sequences 1, 2, 5 and 6, whereas in Seq.3 and 4
it becomes more apparent. Here, the IWF yields the
highest resolution of all, judging from its ability to resolve
the first line of text "A BEAR". In that regard it even
outperforms the single observation PCA algorithm, in quality as
well as in speed. Otherwise, the IBD yields the sharpest
result for Seq. 3 and the LRD the smoothest without being
blurry, unlike the PCA result.
Proc. of SPIE Vol. 7108 71080M-9
-
eMtMt
r
sample image average 100 median 100 average 1000
IBD average IBD median LRD average LRD median
standard deviation 1000 IWF average IWF median PCA single C1234
for average 100
PCA single C1 for average 100 PCA single C2 for average 100 PCA
single C3 for average 100 PCA single C4 for average 100
PCA single C1 for median 20 PCA single C1234 for median 20 PCA
single C1 for median 10 PCA single C1234 for median 10
Proc. of SPIE Vol. 7108 71080M-10
-
PCA multi C1 for 10 PCA multi C1234 for 10 PCA multi C1 for 20
PCA multi C1234 for 20
PCA multi C1 for 100 PCA multi C4 for 100 PCA multi C34 for 100
PCA multi C1234 for 100
Fig. 6 (spreads 2 pages): Comprehensive deconvolution results
from Seq. 7 (1.07 p.m.), recorded in strong turbulence
conditions, beginning with sample and averages as reference,
with standard deviation of total sequence illustrating
turbulence effect on edges, followed by the respective IBD, LRD
and IWF results, each for average and median of 100
frames; next are results from PCA single version applied to
average of 100, median of 20 and 10 frames and last are
results from PCA multi version applied to 10, 20 and 100
frames
For all other sequences, including Seq. 7, IBD and LRD look
essentially the same. This, in conjunction with the passable
IWF results, means that at least our PSF estimates are quite
satisfactory. As regards the noise parameter, we could
discern neither improvement nor worsening in the results
compared to results created under the zero noise assumption.
The PCA single-observation algorithm has proven its ability to
resolve vertical and horizontal lines. Even in Seq. 7 it
manages to distinctly resolve the vertical lines. The first
principal component, corresponding to the greatest variance,
yields the best result. In Fig. 6 (previous page) all four
components are depicted individually in order to illustrate
their
respective characteristics and the false 3D-effect in
particular, which is inherent to the first two components and
obviously increases along with filter support which again
increases with growing turbulence strength. (It should be
mentioned that the black borders around the results of the
single observation PCA version are caused by the algorithm
and depend on the size of the corresponding filter support.)
As to the multi-observation PCA version, for 100 input frames it
fails completely, and increasingly so in accordance with
growing turbulence strength. Concerning its performance for only
10 or 20 input frames, the results for Seq. 7 indicate
that it does not meet either the standard established by the
single version nor that of the other deconvolution algorithms,
or even of the simple average or median operation.
5.1.3 Image Quality Metrics
For lack of a better alternative we used the averages and median
and also the input frame as reference for metric IQM1 in
Eq. (2.1), meaning the results were correlated to the
deconvolution input (i. e. average or median).
As was to be expected, the values for IBD and LRD were almost
identical for both metrics; those for the IWF from
metric IQM2 Eq. (2.2) were generally a bit better but ranged in
the same order as the other two. Due to the (at times
extreme) contrast enhancement of the PCA algorithms, both
versions received high quality marks from both metrics, not
always justified.
On the whole, the results were rather consistent with our
initial assessment but especially the "high quality" of the
multiple observation PCA results for 100 frames signifies that
deconvolution results are prone to noise amplification
resulting in unnaturally high frequency components and high
contrast, thus rendering both metrics somewhat unreliable.
Proc. of SPIE Vol. 7108 71080M-11
-
6. COCLUSIOS
We have described and discussed the various steps in our
algorithm for the restoration of images that were degraded by
atmospheric turbulence. This includes a comparative analysis of
four different deconvolution methods and a presentation
of the respective results for a number of real data image
sequences, both long and short exposure. The partially
automated PSF estimation has produced satisfactory results. The
PCA approach has yielded interesting results and
proven worthy of being pursued further. The general feasibility
of the real-time objective for our project has been
confirmed even if not quite as of yet.
Future work includes replacing the simple Gaussian filter by a
more accurate estimate for the average atmospheric PSF
and finding an analytical description for the correspondence
between Cn and edge width. We also intend to apply MCA locally, by
using block matching, in order to improve compensation for image
dancing in anisoplanatic conditions.
Furthermore, the use of thresholded edge images rather than
greyscale images might help improve the reliability of the
quality metrics.
It should also be mentioned that the full stabilizing effect of
the restoring algorithm(s) can best be demonstrated when
directly comparing the original to the restored images in a
video sequence.
ACKOWLEDGMETS
The authors would like to thank their colleagues, Gabriele
Marchi, Endre Repasi and Wolfgang Schuberth, for supplying last
minute measurements, helpful insights and an abundance of
literature on atmospheric transfer functions.
REFERECES
[1] Vladimir I. Polejaev, Pierre R. Barbier, Gary W. Carhart,
Mark L. Plett, David W. Rush, Mikhail A. Vorontsov,
Adaptive compensation of dynamic wavefront aberrations based on
blind optimization technique, Proc. of SPIE, Vol. 3760, pp. 88-95,
(1999).
[2] G. Marchi, R. Wei, "Evaluation and progress in the
development of an adaptive optics system for ground object
observation", Proc. of SPIE, Vol. 6747, (2007) [3]
E. Mauer, "Investigation of atmospheric turbulence effects on
extended objects in high-speed image sequences applying automatic
image analysis", Proc. of SPIE, Vol. 5237, pp. 39-48, (2004)
[4] E. Repasi, "Image Catalogue of Video Sequences recorded by
FGAN-FOM during the RTG-40 Field Trial",
Distributed to group members, (2006) [5]
J. Gilles, "Restoration algorithms and system performance
evaluation for active imagers", Proc. of SPIE, Vol. 6739, 6739B,
(2007)
[6] Z. Wang, A. C. Bovik, "A universal image quality index",
IEEE Signal Processing Letters, vol. 9, no. 3, pp. 8184,
(2002). [7]
M. A. Vorontsov, G. W. Carhart, D. V. Pruidze, J. C. Ricklin, D.
G. Voelz, "Image quality criteria for an adaptive imaging system
based on statistical analysis of the speckle field", J. Soc. Am. A,
Vol. 13, No. 7, (1996)
[8] D. Kundur, D. Hatzinakos, "Blind Image Deconvolution", IEE
Signal Processing Magazine, 1053-5888/96, pp. 43-
64, (1996) [9]
R. C. Gonzalez, R. E. Woods, "Digital Image Processing", 2nd
ed., Prentice Hall, NJ, (2002) [10]
W. H. Richardson, "Bayesian-Based Iterative Method of Image
Restoration", J. Opt. Soc. Am. 62 (1), pp. 55-60, (1972)
[11] L. Lucy, "An iterative technique for the rectification of
observed distributions", Astron. J. 79, pp. 745, (1974)
[12] Ayers G. R., and Dainty J. C., Iterative blind
deconvolution method and its applications, Opt. Letters, vol. 13,
no.
7, pp. 547549, (1988). [13]
D. Li, M. Mersereau, S. Simke, "Atmospheric Turbulence-Degraded
Image Restoration Using Principal Component Analysis", IEEE
Geoscience and Remote Sensing Letters, Vol. 4, No. 3, pp. 340-344,
(2007)
[14] M. Greco, C. S. Huebner, G. Marchi, "Quantitative
performance evaluation of a blurring restoration algorithm
based
on principal component analysis", to be published on Proc. SPIE
of Optics in Atmospheric Propagation and Adaptive Systems XI,
(2008)
[15] E. Repasi, R. Weiss, "Analysis of Image Distortions by
Atmospheric Turbulence and Computer Simulation of
Turbulence Effects", Proc. of SPIE, Vol. 6941,69410S, (2008)
Proc. of SPIE Vol. 7108 71080M-12