Preprint

UCRL-JC-139877

Fast Fourier and Wavelet

Transforms for Wavefront
Reconstruction in
Adaptive Optics

F. U. Dowla, J. M. Brase, and S. S. Olivier

This article was submitted to

45th Annual Meeting of the International Society for Optical
Engineering
San Diego, California
July 30 – August 4, 2000
U.S. Department of Energy

Lawrence
Livermore
National
July 28, 2000
Laboratory

Approved for public release; further dissemination unlimited

 DISCLAIMER

This document was prepared as an account of work sponsored by an agency of the United States
Government. Neither the United States Government nor the University of California nor any of their
employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for
the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or
represents that its use would not infringe privately owned rights. Reference herein to any specific
commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, does not
necessarily constitute or imply its endorsement, recommendation, or favoring by the United States
Government or the University of California. The views and opinions of authors expressed herein do not
necessarily state or reflect those of the United States Government or the University of California, and
shall not be used for advertising or product endorsement purposes.

This is a preprint of a paper intended for publication in a journal or proceedings. Since changes may be
made before publication, this preprint is made available with the understanding that it will not be cited
or reproduced without the permission of the author.

This report has been reproduced

directly from the best available copy.

Available to DOE and DOE contractors from the

Office of Scientific and Technical Information
P.O. Box 62, Oak Ridge, TN 37831
Prices available from (423) 576-8401
http://apollo.osti.gov/bridge/

Available to the public from the

National Technical Information Service
U.S. Department of Commerce
5285 Port Royal Rd.,
Springfield, VA 22161
http://www.ntis.gov/

OR

Lawrence Livermore National Laboratory

Technical Information Department’s Digital Library
http://www.llnl.gov/tid/Library.html
 Fast Fourier and Wavelet Transforms for Wavefront Reconstructionin
Adaptive Optics
Farid U. Dowla, Jim M. Brase, and Scot S. Olivier

Lawrence LivermoreNational Laboratory

Livermore, California94551

ABSTRACT

Wavefront reconstruction techniquesusing the least-squares estimators are computationally quite

expensive. We compare waveletandFouriertransformstechniquesinaddressing the computation
issues of wavefront reconstruction in adaptive optics. It is shown that because the Fourier approach is
not simply a numerical approximation technique unlike the waveletmethod, the Fourier approach might
haveadvantagesinterms of numericalaccuracy.However,strictly from a numericalcornputations
viewpoint, the wavelet approximation method might have advantagein terms of speed. To optimize the
wavelet method, a statistical study might be necessary to use the best basis functions or “approximation
tree.”

Key words: Novel algorithms and architectures, adaptiveoptics, wavefront reconstruction, fast
algorithms, real-time imaging, wavefrontcontrol, phase reconstruction, FFT methods, wavelets, MEMS.

1. INTRODUCTION

Wavefront reconstruction techniquesfor real-time imaging applications using adaptive optics are
computationally intensive. As we approach implementationof adaptive optics systems with thousands
of actuators employing micro-electromechanical systems (MEMS), the use of efficient high-
performance algorithms for wavefront reconstruction willbe of significant importance. In this study we
have evaluated various fast wavelet and Fourier methods in order determine
to both the speed and
robustness of wavefront reconstruction algorithmsfor real-time imaging in large scale adaptive optics
systems.

We developed fast signal processing algorithmsfor wavefront reconstruction. In particular, we

have studied both fastdiscrete wavelet (DWT) and fast Fourier methods (FFT) to determine the
performance of these algorithms in wavefront phasereconstruction. The wavelet method is of particular
interest becausethe order of the computationfor the 1-D DWT i.s onlyO(N) as opposed to O(N Log N)
for the I-I> FFT. The usefulness of the wavelets in terms of reducednumber of computations is
enhanced by the fact that the transform coefficients canbe truncated significantly; i.e. transformed into
sparse expansions. This energy compaction propertyof the 2-D wavelet transforms can be exploited
fully in the wavefront reconstruction algorithm in adaptive optics. Since there are many “good
wavelets” or basis functions, in this study we compare the performance for wavelets for fast and robust
wavefront reconstruction.

In section 2 we first derive the Fourier approach andshow that this approach is really an analytical
alternative to the truesolution. On the other hand, in section 3 we show that the wavelet method is
strictly a numerical approximation to the least square technique.We illustrate our results with the
wavelet scheme, andsummarize our conclusionsin section 4.

2. WAVEFRONT RECONSTRUCTION USING THE F’FT

If we sample a wavefront on a regular 2D grid we have an arrayof phase valuesqPmn

where
0 Im 5 M - 1 and 0 In 5 N - 1. The phase differencesA:,, and A:,, are then given by

We can writethe phase differences as a matrix equation

i=A@

where is a column vector containing the x and y phase differences, is a column vector containing
the phase values, andA is a finite difference matrix. Theleast-squares solution for 3, given a slope
-+

measurement vector A , is given by solving the normalequations

A ~ A= @~~i
In most adaptive optics systems the normal equationsare solved directly using singular-value
decomposition to give the phase estimate

@=( A ~ A ) ” A ~ A

This solution requires(MN>’ operations to implement the real-time phase reconstructor, assuming that
the matrix (ATA)-’A’ is pre-computed.

For large-scale systems with thousandsof phase sample points we need a more computationally efficient
approach, The normal equations can be written in the form [3]:

represented by its discrete Fourier coefficientsas

Now we can write the normal equations interms of the Fourier coefficients as

where ij,,, are the Fourier coefficients of p,,, . For spatial frequency p,q we can solve directly for qp,q
as

Now we need to write ij,,, in terms of the measurements Am,n.

-i2np -
where 3 { f } is the Fourier transform of f . Since 3{AZ+l,n}= e A>.4 we have

Then we can write the Fourier coefficients of the phase estimate-asthe sum of two spatial filter
operations

@P,P = H;,$p, +H ; , , ~ ; , ,

where the filters are given by

The algorithm for the Fourier wavefront reconstructor is given by

1. Compute the Fourier transforms of the phase d8erencesfrom the wavefront sensor, pp,q,
A:,q,
2. Apply the spatial filters Hp”.4,Hp’,qtu get @p,q

3. Inverse Fourier transform tu get v,,,,~

If N is a power of two the spatial filter operations can be implemented withFFT’s. The computational
requirements then scale as O ( N log, N ) rather than as O ( N 4 ) in the direct vector-matrix multiplication
approach.

3, WAVEFRONT RECONSTRUCTION USING WAVELETS

Although waveletsare well known for their signal and image compression properties [2,5,6],a less well-
known use of the wavelet transformis.in obtaining a fast approximate numerical solution [3] for a
system of linear equations. As discussed in the previous section, this implies waveletscan be used in the
adaptive optics wavefront reconstruction problem.The matrix operator, in solving a system of linear
equations, can be thought of as a two-dimensional image. Taking advantage of the energy compaction
property of the two-dimensional wavelettransform, we can expect a large fraction of the wavelet
coefficients to be small and negligible. Hence, the linear system to be solved, in the transform domain,
is a sparse system (mostly zero coefficients). To be more specific, consider solving a set of linear
equations of the forrn:

Ax=b
The two-dimensional wavelet transformof the system matrix, A, and the vector b are denoted by:

A=WAWT

b=Wb

where W represents the kernel of the wavelet transform operator. Now we can solve for
where 2 represents the wavelet transform of the solution vector. To obtain the solution x, we take a
final inverse of the orthogonal wavelet transform:

x=WT x

The efficiency of the wavelet transformin solving a set of linear equations will ultimately depend both
on the sparseness of the matrix and on the particular wavelet transform approximation algorithm. For
example, consider the 2-D wavelet transform depicted in Figure 1.

-b LP-cols "w

W LP-rows

3
Coefficients
+ LP-cols -+
b HP-~ows '

"FL .HP-cols +
u
Figure 1 : A block diagram of the forward wavelet transform for two-dimensional waveforms or images. Note that for the
inverse transform, simply replace downsampling (down-arrows) by upsampling (up-arrows) and right-arrows by left arrows.

The accuracy of the numericalapproximation and the numberof computation will depend on the
number of decomposition steps and will further depend on the filterbanks "tree" or path employedfor
the transformation [5].There are also a number of different approximation trees that can be chosen
from; i.e. the exact path of lowpass and highpass filters. We studied lowpass, highpass, and both
lowpass and highpass approximationsfor adaptive optics wavefront reconstruction. In our
implementation scheme,we found that the lowpass wavelet coefficients were generally more useful.
The results in this paper are based on thedouble lowpass wavelet approximation coefficients. The
inverse wavelet transformis very similar to the forward transform shown in Figure 1, except that the
down-sampling by two (down-arrows) are replaced by up-sampling (placing zeros in between data
samples) and the right-arrowsare to be replaced by left arrows. Finally, the set of filter coefficients used
to compute the wavelet transforms and the inverse wavelet transforms is shown in Figure 2.
 Waelet Coefficients [Lo-A, Hi-A, Lo-S, Lo-A]
I r I I I I I I 1

t-

4
1
-1
- I I I I
I I

1 2 3 4 5 6 7 0
I

01

-1
3
'
r I

I
I

I
b

I
1

I,

I
I

II
L

I
I

I
3

1
1 2 3 4 5 6 7 8
II 1 I I I I I I

-1 I
I I I I I

1 2 3 4 5 6 7 8
11 I I I I I 1 I

-1 '
1 2
I

3
I I

4 5
I I

6
I

7
I
8

Figure 2: Plots of the typical wavelettransform filter coefficients:lowpass analysis, highpass analysis, lowpass
reconstruction, and highpass resconstruction filters.

In summary, the algorithm can be described as follows:

Step 1: Compute the wavelet matrix approximationfor a single stage wavelet transform. This step is
precomputed just once, and dependson the system.

Step 2: This step is repeated multiple times:

2.1 Compute the wavelet transform of the input vector;
2.2 Compute multiplication in the wavelet domain;
2.3 Reconstruct the vector approximation by taking the inverse wavelet transform.

Note that becausethe wavelet transform basis has only a small number of terms (localized) and because
of the downsamplingoperations, it can be shown that the computation of the wavelet transform is O(M)
in 1-D and O( M *) in 2-D, where M is the dimension of the input. The computational requirements for
the wavelet wavefront reconstruction approach thenscales as 0 ( N 2 ) rather than as O(iV4) in the direct
vector-matrix multiplyapproach, and O(N log, N ) using the FFT method. A note of caution: although
the wavelet transform is numerically most attractive, it is important to point out that this approachis an
approximation and its usefulness will ultimately dependon how well it does in terms of numerical
accuracy. In the next section we presentresults from applying the wavelet reconstruction method of real
wavefront data.
 4. NUMERICAL EXAMPLES USING WAVELETS

In section 4.1 we first illustrate the wavelet approximation schemeusing a simple tractable numerical
example. Performance of the approximation on real data is illustrated in sections 4.2 and 4.3. Finally,
comments on areas of care required for using this methodare included.

4.1 A SIMPLE NUMERICAL ILLUSTRATION

For the sake of clarity we first consider simple numerical example on matrix-vector multiplication using
the wavelet transfonn scheme. Consider a (8x8) matrix A, and a (8x1) column vector x shown below:
”

64 2 3 6 1 6 0 6 7 57 1
9 55 54 12 13 51 50 16 2
17 47 46 20 21 43 42 24 3
40 26 27 37 36 30 31 33 4
A= ,x=
32 34 35 29 28 38 39 25 5
4i 23 22 44 45 19 18 48 6
49 15 14 52 53 11 10 56 7
8 58 59 5 62 63
4 1
- 8.

The numerical solutions to the direct matrix-vectormultiplication and the wavelet approximation are
respectively given by

T
Ax=[1162, 1178, 1178, 1162,1162, 1178,1178, 11621

w”[A x ] = [I 170,1170,1170,I170,1170,I170,1170,1
170IT

Note that in the waveletapproximation method, we used the 2ndorder Daubechies wavelet coefficients
with two consecutive lowpass (or smooth approximation)filtering and downsampling of the rows and
the columns. The solution therefore represents a smooth approximation of the exact solution. In
general, thisis an important point about the wavelet approximation method on real data: the
approximation qualitydepends on the exact decompositiontree (Le. the path of lowpass and highpass
approximations) andon determining how well a particular tree is able to represent the matrix A and
therefore also the solution vector A x.

4.2 EXAMPLES WITH REAL DATA

An example of the A 0 system matrix is shown as an image in Figure 3. To illustrate the

approximation with a discernable plot, weshow part of the reconstruction result in Figure 4, and
compare this wavelet approximation solution withthe conventional reconstruction technique. In this
particular example, we used a linear ramp signal forx, the slope measurement vector.

50

100

150

200

250

300

350

400

450

50 100 150 200 250 300

Figure 3: The A 0 System Matrix (A} represented as an image

Although the wavelet transform technique iscomputationally highly efficient, through the course of this
study we discovered that the approximationquality does depend on the form of system matrix, A, and on
the multiplying vector, x, Furthermore, theexact lowpass-highpass combinations used in the
approximation algorithmis also quite important. Our conclusion is that to use the wavelet transform
technique successfully, wewould need to first carefully characterize the statistical properties of the
distorted wavefronts and thenselect the best “basis” (the lowpasshighpass tree). A significant amount
of work has been performed in wavelet packets 161 that could be used for optimal performance. Finally,
we plan to evaluate the performanceof this in a full adaptive optics system simulation mode.
 x lo5 Matrix-bector multiplication with real A 0 system matrix
I I I I I I I I I I I
2.792

2.79

2.788

2.786

2.784 i
2,782

2.78

2.778

2.776
110 I I

20
I

30
I

40
I

50
I

60
I

70
I

80
I

90
I

100
1

110

Figure 4: The matrix-vector multiplicationresult y = Ax, where A is shown in Figure I and x is a linear ramp vector [ 1 2 ...
3491. Only a section of 112 points of the solution vector and the corresponding wavelet approximate smooth solution is
plotted over. The smoother graph represents the wavelet approximation solution.

5. SUMMARY AND CONCLUSIONS

In this paper we discussed how the FFT and the wavelet transform methods could be used to speed
up computation in adaptiveoptics wavefront reconstruction. We were particularly interested in
comparing wavelet method withthe FFT technique. Although, in terms of the number of operations, the
wavelet transform seemsto be moreattractive as discussed in sections 2 and 3, we find that the wavelet
approximation mightor might not be accurateenough. Only a full scale simulation study can determine
the performance in terms numerical accuracy; we plan to do this in the near future. The accuracy of the
approximation dependson the particular vector andmatrix properties, and on the choice of the wavelet
tree used to decompose the tree. A statistical study will be needed to select the optimal wavelet
decomposition. On the other hand, as discussed in section 2, the Fourier approach is much less of a
numerical approximation, andmore of an analytical solution until the application of the FFT. Hence,
the FFT approach mightbe more suitable to minimize loss in the numerical accuracy.

We observed that the type of wavelet basis function used does not seem to make a significant
difference in the numerical results. A low order basis function, like the Daubechies 2ndorder
coefficients, seems to be adequate with real data.
 6. ACKNOWLEDGEMENTS
Work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National
Laboratory under contract W-7405-ENG-48.

7. REFERENCES

1. Anant, K. S., F. U. Dowla, G. H. Rodrigue. Vector quantization of ECG wavelet coefficients,

IEEE Signal Processing Letters, pp. 1-4, July, 1995.

2 . Bowman, K. and W, Rhodes, “Application of wavelets to wavefront reconstruction in optical

systems, S P E Vol. 3126, August 1997, pp. 288-294.

3. Freischlad, K and C. Koliopoulos, “Wavefront reconstructionfrom -Noisyor Difference Data

Using the Discrete Fourier Transform,” SPIE Vol. 55 1, Adaptive Optics 1985, pp. 74-80.

4. Hunt, B. R., “Matrix formulation of the reconstructionof phase values from phase differences,”
J. Opt. SOC. Am., Vol. 69, No. 3, March 1979, pp. 393-399.

5. Vaidyanathan, P. P, “Multirate Systems and Filterbanks,” Prentice-Hall, 1993.

6. Vetterli, M. and J. Kovacevic, “Wavelets and Subband Coding, Prentice-Hall, 1995.