Title: Wavefront Sensing II
1Wavefront Sensing II
- Richard Lane
- Department of Electrical and Computer Engineering
- University of Canterbury
2Contents
- Session 1 Principles
- Session 2 Performances
- Session 3 Wavefront Reconstruction
3Session 2 Performances
- Geometrical wavefront sensing take 2
- The inverse problem
- The astronomical setting
- The basic methods
4Geometric wavefront sensing(or curvature sensing
without curvature)
Plane 1
Image Plane
Plane 2
Improve sensitivity (signal stronger)
Improve the number of modes measurable (signal
weaker)
5Geometric optics model
- Slopes in the wave-front causes the intensity
distribution to be stretched like a rubber sheet
- Aim is to map the distorted
- distribution back to uniform
6Geometric wavefront sensingTake 2
Intensity Plane 1
Plane 1
Image Plane
Intensity Plane 2
Plane 2
Intensity distribution gives the probability
distribution For the photon arrival
7Recovering the phase
Intensity Plane 1
Integrate to Form CDF
Intensity Plane 2
Choose level
Probability density functions
Integrate slope to find the phase
Defocus!
8Forward Problem
9Inverse Problem
Performance is determined by amount of photons
entering the aperture and assumptions about the
object and turbulence
10Imaging a star
11Multiple layers
For wide angle imaging we need to know the
height of the turbulence
Layer 1
h1
h2
Layer 2
Aperture Plane
12The fundamental problem
- How to optimally estimate the optical effects of
turbulence from a minimal set of measurements
13Limiting Factors
- Technological
- CCD read noise
- Design of wavefront sensor (Curvature,
Shack-Hartmann, Phase Diversity) - Fundamental
- Photon Noise
- Loss of information in measurements
- Quality of prior knowledge
14In Its Raw Form the Inverse Problem Is Always
Insoluble
- There are always an infinite number of ways to
explain data. - The problem is to explain the data in the most
reasonable way - Example Shack-Hartmann sensing for estimating
turbulence
15Example fit a curve to known slopes
- Solution requires assumptions on the nature of
the turbulence - Use a limited set of basis functions
- Assume Kolmogorov turbulence or smoothness
16Parameter estimation
- Essentially we need to find a set of unknown
parameters which describe the object and/or
turbulence - The parameters can be in terms of pixels or
coefficients of basis functions - Solution should not be overly sensitive to our
choice of parameters. - Ideally it should be on physical grounds
17Bayesian estimation 101An important problem
- And if you know that it models two people
- splitting the bill in a restaurant?
18Possible phase functionsZernike basis
?
Zernike Polynomials Low orders are smooth
Pixel basis, highest frequency 1/(2?)
19Estimation using Zernike polynomials
phase weighting Zernikie polynomial
- Measurement Interaction Zernike Polynomial
- vector matrix
Coefficents
- ith column of T corresponds to the measurement
that would occur if the phase was the ith Zernike
polynomial
20Extension to many modes
- Provided the set of basis functions is complete,
the answer is independent of the choice - The best functions are approximately given by the
eigenfunctions of the covariance matrix C - These approximate the low order Zernike
polynomials, hence their use. - Conventional approach is to use a least squares
solution -
- and estimate only the first M Zernikes when M
N/2 (N is the number of measurements)
21Ordinary least squares
Weighted least squares
- Not all measurements are equally noisy
- hence minimise
22Conventional Results
- As the number M increases the wavefront error
decreases then increases as M approaches N. - Reason when MN there is no error and there
should be as higher order modes exist and will be
affecting the measurements
23Phase estimation from the centroid
- Tilt and coma both produce displacement of the
centroid - According to Noll for Kolmogorov turbulence
- Variance of the tilt
- Variance of the coma
Ideally you should estimate a small amount of coma
24Bayesian viewpoint
- The problem in the previous slide is that we are
not modelling the problem correctly - Assuming that the higher order modes are zero, is
forcing errors on the lower order modes - Need to estimate the coefficients of all the
modes as random variables
25Example of Bayesian estimation for
underdetermined equations
- Measurement z is a linear function of two
unknowns x,y
- The estimate (denoted by ) is a linear
- function of z
- We want to minimise the expected error
Statistical expectation
26Minimisation of the error
- Key step, rewrite in terms of and
- Solution is a function of the covariance of the
unknown - parameters
27Vector solution for the phase
- Express the phase as a sum of orthogonal basis
functions - Observed measurements are a linear function of
the coefficients - Reconstructor depends on the covariance of a
28Simple example for tilt D/r04
- From Noll
- From Primot et al
29Bayesian estimate of the wavefront
Minimizes
30Summary Bayesian method
- When the data is noisy you need to put more
emphasis on the prior. - For example, if the data is very bad, dont try
and estimate a large number of modes - When done properly the result does not depend
strongly on C being exact - Error predicted to be
- where
31Operation of a Bayesian estimator
- Minimizes
- When D becomes very large, the data is very noisy
then more weight is placed on the prior - data prior
- Ultimately as D?8, a?0 (for very noisy data no
estimate is made)
32Bayesian examination question
- You are on a game show.
- You can select one of three doors
- Behind one door is 10000, behind the others
nothing - After you select a door, the compere then opens
one of the other doors revealing nothing. - You are given the option to change your choice
- Should you?
33Estimating the performance limits when it is
non-Gaussian
- The preceding analysis is fine when the
measurement errors can be modelled as a Gaussian
random variable - On many equations you need to perform an analysis
to work out the error in the analysis - Cramer-Rao bounds
34Cramer-Rao bound
- Linear unbiased estimators only
- Essentially the quality of the parameter estimate
is given by the curvature of the pdf - Doesnt tell you how to achieve the bound
35Simple example
- Find the performance limit estimating the mean of
a one-dimensional Gaussian from 1 sample
36Points to note
- Limit is a lower bound. Clearly for 1 sample from
the pdf it cannot be attained - The variance decays as 1/N with more samples
- For a Gaussian asymptotically the centroid of the
distribution can be shown to approach the
Cramer-Rao bound
37Estimation of a laser guidestar location,
Cramer-Rao bound
Small projection telescope
Large AO corrected projection telescope
Large uncorrected projection telescope
Key points In the presence of saturation a
focused spot may not be optimal Need to know the
pattern to reach the limit
38Optimal estimation of a parameterwavefront tilt
- Important because the wavefront tilt is the
dominant form of phase aberration - A small error in estimating the tilt can be
larger than the full variance of a higher order
aberration.
39Issues
- Displacement of the centroid of an image is
proportional to the average tilt (not the least
mean square) of the phase distortion - Will discuss this issue later, but for the moment
concentrate on estimating the mean square tilt.
40How do you estimate the centre of a spot?
- The performance of the Shack-Hartmann sensor
depends on how well the displacement of the spot
is estimated. - The displacement is usually estimated using the
centroid (center-of-mass) estimator. - This is the optimal estimator for the case where
the spot is Gaussian distributed and the noise is
Poisson.
41Centroid estimation for a sinc2 function
42Why Not Use the Centroid?
- In practice the spot intensity decays as
- This means that photons can still occur at points
quite distant from the centre. - Estimator is divergent unless restricted to a
finite region in the image plane
43Diffraction-limited spot
- For a square aperture, the distribution is
44Photon arrival simulation
45Solutions (1)
- Use a quad cell detector and discard the photons
away from the centre - The signal from the outer cells is discarded
because it adds too much noise
46Solutions 2
- Use an optimal estimator that weights the
information appropriately - Consider two measurements of an unknown parameter
an estimate of a parameter with different
variances - A weighted sum is always a better estimator
- A non linear estimator is better still
47Maximum-likelihood estimation
- If photons are detected at x1, x2, xN, the
estimate is the value that maximizes the
expression - The Cramer-Rao lower bound for the variance is
- For a large number of photons, N, the variance
approaches the Cramer-Rao lower bound.
48Centroid location by model fitting
- Technique relies on finding a model of the object
- Not sensitive to the size of window (unlike the
centroid) - Centroid is a closed form
solution for fitting a Gaussian of variable width
49Tilt estimation in curvature sensing
- The image is displaced by the atmospheric tilt,
how well you can estimate it is determined by the
shape of the image formed.
50Tilt estimation in the curvature actual
propagated wavefronts
51Performance versus detector positionfor a
curvature sensor
52Actual Wavefront sensor data
- Observation at Observatoire de Lyon
- SPID instrument on 1-m telescope
- 20x20 Shack-Hartmann lenslet array
- Exposure time 2ms
- Objects Pollux, point object 2500 frames
- Castor3 arc second binary 2500
frames
53Centroiding issues
- Accuracy required to a fraction of a pixel
- Sampling rate 60 of Nyquist
54Finding the model
- Need a good model of the object
- In each lenslet of the Shack-Hartmann acts like a
small telescope the dominant effect is one of
tilt. - gt We have a large number of images of the same
object shifted before they are sampled.
55Solution approach
- Use blind deconvolution to find model
- MAP framework (Hardie et al, FLIR)
- Data-capturing process
- Choose initially so that
- Prior information
- Laplacian smoothness for the optics
- Maximum entropy for CCD
56Typical SPID data frames
Single Wavefront Sensor Frame
Long term WSF
Blow up of a spot
Movie of a spot
57Simulations
- Inputs
- Object f point source
- Optics diffraction-limited pattern of square
aperture - CCD structure Gaussian-like
- Random displacements
- White Gaussian noise dB, 30dB, 15dB
58Simulation result 15dB noise
Optics reconstruction
CCD reconstruction
59Traditional centroiding
- Centre of gravity of spot image
- Problems
- Finite pixel size
- Finite window size
- Readout noise (more pixels more noise)
- Bias
- Problems become worse with extended objects
60Model-fitting
- Full blind deconvolution computationally
unreasonable - Fit a model estimated by blind deconvolution
- Use model to determine centroids
61Error in centroid calculation
62Blind deconvolution results
Optics reconstruction
CCD reconstruction
63Results from speckle image deconvolution
(narrowband)
Binary estimated with model fitted centroids
Binary estimated with traditional centroids
64Phase reconstructions of Binary
Model based centroiding
Traditional centroiding
65Conclusions
- Bayesian approaches provide a logical framework
for filling in missing data - Make sure of what you are assuming
- Cramer-Rao bound can provide a performance limit
- You need to look at the whole process when
deriving an algorithm
66And the answer is(ref Stark and Woods)
?
?
?
67Actual Wavefront sensor data
- Observation at Observatoire de Lyon
- SPID instrument on 1-m telescope
- 20x20 Shack-Hartmann lenslet array
- Exposure time 2ms
- Objects Pollux, point object 2500 frames
- Castor3 arc second binary 2500
frames
68Subpixel displacement estimation
Wavefront sensing is based on estimating the
tilts produced by atmospheric distortion, the
accuracy of displacement estimation is critical.
Estimated optics psf
Data from SPID 2500 frames undersampled by 40
Spot displacements
Estimated CCD pixel sensitivity
69Explanation of the terms
70Possible phase functionsZernike basis
71The inverse problem
72Alternatively
73Prior information
- Infinite number of unknowns, but a finite number
of centroid measurements from the sensor - Conventional approach is to choose the basis
functions and estimate M coefficients, where M lt
N the number of measurements
74Using real data binary star
- 14x14 Shack-Hartmann lenslet array
- Exposure time 3.2ms
- Object Castor, a binary star
- Intensity ratio 2.1
- Separation 3.1 arcseconds
75Blind deconvolution results
- Intensity ratio 2.4
- Separation 3 arc seconds