Wavefront reconstruction using iterative Fourier transforms Frangois Roddier and Claude Roddier University of Hawaii, Institute for

Astronomy, 2680 Woodlawn Drive, Honolulu, Hawaii 96822. Received 4 February 1991. Sponsored by James B. Breckinridge, Jet Propulsion Lab­ oratory. 0003-6935/91/111325-03$05.00/0. © 1991 Optical Society of America. The problem of obtaining a linear least-squares wavefront estimate from an array of wavefront slope measurements has long been discussed in the literature. 1,2 Noll3 has shown that the solution is that of a Poisson equation with Neumann boundary conditions. Succesive over relaxation (SOR) algo­ rithms are generally used to solve the problem numerically.4 Here we present a novel algorithm based on fast Fourier transform (FFT) subroutines. The basic idea is that the Laplacian operator ▿2 = ∂2/∂x2 + 2 ∂ /∂y2 translates into a multiplication by u2 + υ2 in the Fourier (u,υ) domain. Hence by taking the Fourier trans­ form of the wavefront Laplacian ▿2W, dividing it by u2 + υ2, and taking an inverse Fourier transform one should be able to retrieve the wavefront surface W. This is the basis of Fourier methods for solving differential equations. This simple approach would hold for a wavefront with no bound­ ary (function with unbounded support). In practice the wavefront Laplacian is multiplied by the pupil transmission function Pix,y), and its Fourier transform is, therefore, con­ volved by the Fourier transform of P{x,y). As a result the simple property indicated above no longer holds. A proper treatment should take the signal boundaries into account using boundary conditions. Fourier methods have already

been proposed in the literature,5,6 but they essentially apply to square or rectangular domains. The method described here applies to domains with any arbitrary shape. The basis of the method is an extrapolation of the wavefront beyond the boundaries using a Gershberg type algo­ rithm. 7 A flow chart of the algorithm is shown in Fig, 1. One starts with two arrays of numbers which are the sampled wavefront x and y slopes inside the pupil domain and zeros outside the domain. An FFT algorithm is used to take the Fourier transform of these arrays which represent functions of u and υ. The Fourier transform of the x slope is multiplied by u, that of the y slope is multiplied by υ. Both arrays are added, and the result is divided by u2 + υ2 everywhere except at the origin where the ratio is undetermined. Since we seek the solution which has zero mean, we put zero at the origin. The division by u2 + υ2 can be considered as an apodization. An even stronger apodization can be used at the beginning to accelerate convergence.8 Next, one takes an inverse Fourier transform. A result of the apodization is that the signal now extends beyond the initial boundaries. However, we are interested only in the part within the boundaries which is our first wavefront estimate. The next step consists of taking the x and y derivative of this estimate. The result will in general differ from the original data. The rms difference between the two set of numbers is a measure of the current error in the reconstruction process. The algorithm consists of putting back the original data inside the domain where measurements have been made, whereas keeping the extrap­ olated signal outside the domain. The process is then iterat­ ed. When the error becomes below a preassigned level, the next wavefront estimate is computed and the iterative pro­ cess is stopped. Both the above described algorithm and a SOR algorithm have been applied to the same simulated data. On 128 X 128-pixel arrays the computing time was found to be approx10 April 1991 / Vol. 30, No. 11 / APPLIED OPTICS 1325

Fig. 1. Flow chart of the iterative Fourier transform algorithm used to reconstruct a wavefront W from the measured wavefront slopes ∂W/∂x and ∂W/∂y. imately the same. However, the new algorithm was found to produce better wavefront estimates especially near the edges. This is clearly demonstrated in Fig. 2, which shows an example of wavefront reconstruction. The full line is a section of the original 2-D wavefront (spherical aberration). A circular pupil was used with a 30-pixel radius. The dashed line is the same section reconstructed with the SOR algo­ rithm. The dash/dot line is the section reconstructed with the FFT based algorithm. The approach described above is quite general and can also be used to reconstruct wavefronts from wavefront La-

Fig. 3. Flow chart of an iterative Fourier transform algorithm used to reconstruct a wavefront W from the wavefront Laplacian mea­ sured from out of focus images.10 placians obtained directly from extra-focal images.9,10 Fig­ ure 3 shows a flow chart of a similar algorithm which recon­ structs wavefronts from wavefront Laplacians estimated by taking the difference between the illuminations in symmetri­ cally defocused images. The boundary conditions are taken into account by forcing to zero the radial derivative within a narrow band surrounding the boundaries. This is equiva­ lent to solving the Poisson equation with a zero radial edge slope as a Neumann boundary condition. Indeed the differ-

Fig. 2. Example of wavefront reconstructions from wavefront slopes. Full line, original wavefront; dashed line, reconstructed wavefront using the SOR algorithm. Dash/dot line, reconstructed wavefront using iterative Fourier transforms.
1326 APPLIED OPTICS / Vol. 30, No. 11 / 10 April 1991

ence between extra-focal images produces a narrow edge signal which is a measure of the wavefront radial edge slope. This signal can also be considered as a Laplacian equal to the difference between the wavefront radial slopes on each sides of the edge. By forcing the outer slope to zero, one forces the inner slope to be equal to the edge signal. Unlike previously described algorithms, there is no need to somewhat arbitrarily isolate the boundary wavefront slope signal from the inner wavefront curvature signal. Here one takes the sensor signal as a whole. Clearly the Gershberg algorithm can be similarly used to solve numerically any linear differential equation with boundary conditions on any contour shape using FFT subroutines. To our knowledge, this very interesting application of the Gershberg algorithm has not been considered before. This work has been supported by a grant from the Strategic Defense Initiative Organization, Office of Innovative Science and Technology, and managed by the Harry Diamond Laboratories. References
1. D. L. Fried, "Least-Square Fitting a Wave-Front Distortion Estimate to an Array of Phase-Difference Measurements," J. Opt. Soc. Am. 67, 370-375 (1977).

2. R. H. Hudgin, "Wave-Front Reconstruction for Compensated Imaging," J. Opt. Soc. Am. 67, 375-378 (1977). 3. R. J. Noll, "Phase Estimates from Slope-Type Wave-Front Sensors," J. Opt. Soc. Am. 68, 139-140 (1978). 4. W. H. Southwell, "Wave-Front Estimation from Wave-Front Slope Measurements," J. Opt. Soc. Am. 70, 998-1006 (1980). 5. R. L. Frost, C. K. Rushforth, and B. S. Baxter, "Fast FFT-Based Algorithm for Phase Estimation in Speckle Imaging," Appl. Opt. 18, 2056-2061 (1979). 6. K. R. Freischald and C. Koliopoulos, "Modal Estimation of a Wave Front from Difference Measurements Using the Discrete Fourier Transform," J. Opt. Soc. Am. A 3, 1852-1861 (1986). 7. R. W. Gershberg, "Super-Resolution Through Error Energy Reduction," Opt. Acta 21, 709-720 (1974). 8. L. M. Kani and J. C. Dainty, "Super-Resolution Using the Gershberg Algorithm," Opt. Commun. 68, 11-17 (1988). 9. F. Roddier, "Curvature Sensing and Compensation: A New Concept in Adaptive Optics," Appl. Opt. 27, 1223-1225 (1988). 10. F. Roddier, C. Roddier, and N. Roddier, "Curvature Sensing: A New Wavefront Sensing Method," Proc. Soc. Photo-Opt. Instrum. Eng. 976, 203-209 (1988).

10April 1991 / Vol. 30, No. 11 / APPLIED OPTICS

1327

Sign up to vote on this title
UsefulNot useful