V Multiple Scattering in HPIV: Use of ODT Analysis Techniques J Lobera and J M Coupland Wolfson School of Mechanical and Manufacturing Engineering Loughborough University, Ashby Road, Loughborough, Leics. LE11 3TU J.Lobera@lboro.ac.uk Abstract. Holographic Particle Image Velocimetry (HPIV) has been used successfully to make three-dimensional, three-component flow measurements from holographic recordings of seeded fluid. It is clear that measurements can only be made in regions that contain particles, but simply adding seeding results in poor quality images that suffer from the effects of multiple scattering. Optical Diffraction Tomography (ODT) provides a means to reconstruct a 3D map of refractive index from coherent recordings of scattered fields with different illumination conditions. In this paper we consider from first principles, how these methods can be used to extract information from HPIV recordings in the presence of multiple scattering. We discuss 3D reconstruction using the Born approximation and the Born Iteration Methods. Finally we consider the application of some optimization methods that have been proposed elsewhere as a means to solve strong scattering problems, but simplified with the use of the a-priori information concerning the particle shape and refractive index. 1. Introduction Holographic Particle Image Velocimetry (HPIV) provides a means to make simultaneous three component, three-dimensional (3C3D) measurements of a seeded fluid flow[1],[2]. Implicit in the analysis of HPIV recordings is the assumption that light scattered from a laser source is recorded directly by the hologram such that multiple scattering is negligible. In practice, however, multiple scattering effects appear to increase background noise, thereby decreasing the SNR and ultimately limiting the number of velocity vectors that can be retrieved from a given flow field [3]. Optical Diffraction Tomography (ODT) provides a means to reconstruct a 3D map of refractive index from coherent recordings of scattered fields with different illumination conditions. In 1969 Wolf proposed a method applicable to weak scattering objects, based on the first Born Approximation[4], and more recently the Born Iteration Method (BIM), the distorted Born Iteration Method (dBIM)[5] and other optimization strategies[6] have been proposed as methods to solve stronger scattering. It is clear that the ability to identify particle images, even in the presence of partial obscuration and multiple scattering, would be beneficial to provide detailed velocity measurements from HPIV recordings. It is worth noting however, that ODT methods generally require significant computational effort and only the very simplest of the techniques are currently viable in practical HPIV studies. Nevertheless the methods of ODT are worthy of consideration and in what follows we discuss the techniques’ capability for the case of a simplified problem to reduce the computational overhead. Consider a holographic reconstruction of two particles, back illuminated with a plane wave such that the light scattered from one particle illuminates the other and the second particle effectively obscures the first. To simplify matters, we consider a planar (2D) system with polarization such that all E-fields (i.e. illuminating and scattered fields) are out of the plane and together are solutions to the Helmholtz equation (see Section 2). To investigate the performance of this methods we have developed computational tools to solve the Helmholtz equation for a known refractive index distribution and illuminating field – the forward problem. The solver runs in Matlab and uses the Finite Element Method (FEM) with a rectangular mesh. An important aspect of the solver is the use of an Absorbing Boundary Condition (ABC) or Perfect Matched Layer that surrounds the computational domain[7]. This type of boundary condition allows outgoing waves to pass freely from the computational domain without reflections. A rectangular mesh, with a spacing of one tenth of the wavelength is used. Figure 1a) and 1b) show the scattered fields that correspond to particles of 0.05m and 1m respectively, and are separated by distance of 5m. The particles have a refractive index 1.33 and are assumed to be suspended in air. It can be seen that the smaller particles scatter in a near iso-tropic manner (i.e. as point sources), while forward scatter dominates for larger particles. Let us suppose that the cone of forward scattered light is recorded holographically from these systems and reconstructed in the usual way. Figure 2a) and 2b) show the intensity of the reconstructed images for the case of 0.05m and 1m particles when the recording is made with a large numerical aperture (NA=0.85). For the case of the 0.05m particles, it can be seen that the recording NA is sufficient to resolve the particles in the depth direction. For the case of the 1m particles, it can be seen that although the recording has sufficient depth resolution the images are confused. Furthermore, figure 2c) shows the intensity of the reconstructed images for the case of 1m particles when the recording is made with a smaller numerical aperture (NA=0.3) and it can be seen that the two particles appear as one. a) b) Figure 1 Real part of the scattered field form two particles: a) of 0.05m diameter, b) of 1m diameter. a) b) c) Figure 2 Reconstruction image of two particles: a) 0.05m diameter particles recorded in a NA=0.85 hologram; b) 1m diameter particles recorded in a NA=0.85 hologram; c) 1m diameter particles recorded in a NA=0.3 hologram. It is clear from these results that the recording NA plays a very significant role in HPIV recording and in the past we have made use of relatively large holographic plates, placed close to the object of interest to optimize this parameter[8]. With a view to digital recording, large NA is only practically achievable for the case of small scale flows due to space-bandwidth limitations[9], and for this reason many researchers have considered reconstruction from simultaneous, low NA recordings made from at least 2 different directions. The recordings can be made incoherently, as in stereo PIV (with a thick light sheet)[10],[11] or coherently[12]. In the incoherent case all focus information is lost and a small aperture is used to maintain a reasonable depth of field. In the coherent case some focus information is retained although for a given sized object, the NA is ultimately limited by the camera resolution. In the (coherent) digital holographic method reported by Sheng et al.[12], a forward scattering geometry using two different illuminating waves (at 90 degrees) was used and intensity images were computed from each hologram. Using our 2D system of 2 particles (Fig. 3a), an image of this kind has been computed and is shown in figure 3b). For comparison, we have used an NA=0.3 (as in figure 2c) and it is clear that now the particle positions can be deduced from the brightest peaks in the image. If the particles are repositioned slightly (turned through 45 degrees as in figure 3c), however, then this type of reconstruction gives rise to the distribution shown in figure 3d). Here we note that additional “ghost particles” appear in the reconstruction. Elsinga et al.[11] have shown that the generation of ghost particles is reduced significantly if recordings at more angles are combined and furthermore enhanced by a reconstruction and measurement algorithm that they call Multiplicative Algebraic Reconstruction Tomography (MART). In essence, MART pre-processes each image to decide if, and where, particles exist and then calculates epipolar lines in a manner similar to classical photogrammetry[13]. The crosscorrelation of the epipolar intersection points identified in a given region of 3D space is then calculated to estimate particle displacement. a) b) c) d) Figure 3 Two in-line holograms with NA=0.3: a) two particles aligned with one observations directions; b) reconstructed image of a); two particles rotated 45 degrees, d) reconstructed image of c). Although, some impressive results have been reported, we note that the MART algorithm is highly non-linear both in the particle identification phase and photogrammetric reconstruction and consequently it is expected to be intolerant to optical aberrations and background noise. With HPIV, we have found in the past, that there is significant advantage to collating information in a linear manner (correlation of the complex amplitude) and postponing the non-linear decision process (in this case, deciding which is the signal peak)[14]. Although, not always a linear process, ODT provides a basis to reconstruct the scattering potential (refractive index) from simultaneous recordings of scattered light with different illumination and viewing directions. In the following we outline the basis of ODT and discuss its merit for 3D3C digital HPIV. 2. The scattering problem According to scalar diffraction theory the (complex) amplitude of a monochromatic electric field, E(r), propagating in a medium of (complex) refractive index, n(r), obeys the Helmholtz equation, (1) 2 E(r) k 02 n 2 (r)E(r) 0 where k0 is the free space wave number defined here such that k0=2 , where is the free space wavelength. The electric field can be written as the superposition of, Er(r), the illuminating field (i.e. that which would be present if the medium was absent) and, Es(r), the scattered field. Defining the scattering potential or object function as (r) k 02 n 2 (r) 1 and noting that 2 k 02 E r (r) 0 we have, (2) 2 k 02 E s (r) (r) E r (r) E s (r) The terms on the right-hand-side of equation 2 can be identified as source terms and can be integrated to give (3) s r s 3 E (r ' ) G 0 (r 'r )(r ) E (r ) E (r ) dr where G 0 (r) e jk0 r r is the free-space Green’s Function. It is noted that a holographic microscope (or any far-field optical instrument) is only capable of measuring the propagating part of the scattered field that is observed at a distance from any inhomogeneity. In this, r’>>r and the Green’s Function becomes, (4) exp jk 0 r'r e jk0 r ' r.r' A exp jk.r exp - jk 0 r' where A is a complex constant and k is the wave vector defined as k k 0 r' r' . We can therefore write G 0 (r'-r) r'r r' our observation as, Em(k), given by, E m (k ) A exp( jk .r )(r ) E r (r ) E s (r ) dr 3 (5) Equation 5 can be recognized as the three-dimensional Fourier Transform of the source terms of equation 2 and is general in the sense that it is based only on the assumption of scalar diffraction theory. Although Em(k) appears to be the full spectrum of the source terms, it is noted however, that the Fourier transform is only evaluated over a limited portion of k-space, the (Ewald) sphere, defined by k 2 / . Consequently, the problem is generally under determined and the scattering potential can only be estimated subject to additional properties (for example minimum variance) or certain assumptions as follows. 3. First Born approximation Reconstruction under the assumption of weak scattering was first considered by Wolf 8 in 1969. Let us assume that we use digital holography to make a finite set of measurements E mij, of the monochromatic plane wave components identified by the wave vectors, kj ,that are scattered by a system which is illuminated by a set of plane waves identified by ki . Further assume that in each case the illuminating field, Er(r) is assumed to be significantly stronger than the scattered field, Es(r), such that from equation 5 we have, (6) m 3 E ij A exp( j(k i k j ).r )(r ) dr In this case the measured field values can be recognised as an incomplete set of Fourier coefficients. Accordingly, an estimate of the scattering potential under the assumption of weak scattering, (r) , is given by, 1 (7) m ( r ) A E ij exp( j(k i k j ).r) ij It can be seen that according to the first Born approximation, the object function is a coherent sum of periodic variations in refractive index that are analogous to Bragg gratings formed in a thick (volume) hologram. Returning to the two particle problem of figure 3c), an estimate of the refractive index distribution obtained by applying equation 7 to the FEM data is shown in figure 4a). This result can be compared directly with the image shown in figure 3d) and a significant contrast enhancement can be seen. The improvement can be attributed to the fact that the reconstruction according to the Born approximation is a coherent addition of the information obtained from different views whereas the intensity based reconstruction of figure 3d) is incoherent. We also remark that the Born reconstruction is linear and presents no extra computational burden compared to the incoherent method. Nevertheless it can be seen that the two problematic ghost images remain. 3. Born expansion method Rather than neglect the scattered field, Es(r), in equation 5 Chew and Wang[5] have considered iterative solutions. In this paper we consider the distorted Born Iteration method (dBIM) that essentially use the best estimate of the refractive index, n, as the background object. If we define Er as the solution of 2 E r (r) k 02 n 2 (r )E r (r ) 0 , we can obtain a equation similar to equation 2: (8) 2 k 2 n 2 (r ) E s (r) (r) E r (r) E s (r) 0 1 1 1 Let us remark that in an integral notation the Green function would be not longer a free-space Green function, and that in each iteration, we obtain the increment of the scattering potential 1 (r ) . It is clear that if we start from a background that is air, the solution after the first iteration is the first Born approximation (Fig. 4a). The main drawback of this method appears to be stability and there is not guarantee of convergence. For the case considered before of two particles of water (n=1.33) in air the image obtained is clearly wrong after 3 iterations (Fig. 4b). A similar but weak scattering case of particles with a refractive index n=1.05 is shown in figure 4c) and it can be clearly seen that the contrast has increased. This limitation makes the method unpractical in most cases of interest in HPIV. 4. Optimization method Finally we turn our attention to reconstruction methods that minimise the error defined by J ( ) i E im E ith E im 2 (9) 2 where Eith() is the scattered field obtained by a forward solver. This is a non-linear optimization problem that can be solved by the Conjugated Gradient Method[6]. This method can be applied to strong scattering problems, but the computational cost make unpractical for PIV. In PIV the object is a disperse medium, composed by particles of a known diameter and refractive index value in most of the cases. The use of this a priori information can simplify and reduce significantly the computing task. Returning to the solution obtained from the application of the first Born approximation to the 2 particle problem (Fig 4a), if the 4 possible particle locations are identified we can test whether each is likely to be a ghost particle. In essence, we can test whether removal of particles reduces the discrepancy between measured and predicted results and therefore reduces the error defined by equation 9. In this case, we find the error from our first guess (all 4 particles present) is 0.2989. If we test each particle in turn we find that removal of the bottom-left (ghost) particle has the most significant effect on the error and reduces the value to 0.1244. Removal of top-right the (ghost) particle reduces the error to 0.0065 and note that no further reduction is possible. We conclude that we have the particle distribution shown in figure 4d) and this appears consistent with figure 3c) We note however, that even in this noiseless case the error is not exactly zero and on close examination note that the position of the particles in 4d) are shifted of 0.05m, showing that our initial estimates of particle position were inaccurate. Clearly further dithering of the particle positions could be implemented to decrease the error further. a) b) c) d) Figure 4 Reconstruction image of two particles using: a) 1st Born approximation; b) 3rd iteration of dBIM for particles of n=1.33; c) 3rd iteration for particles of n=1.05; d) optimization method. 5. Conclusions In this paper we have considered the application of the methods of Optical Diffraction Tomography (ODT) to the analysis of HPIV recordings. We used the case of 2 closely spaced particles modeled in 2D to illustrate the effect of multiple scattering and have used the Finite Element Method to compute the scattered fields. From typical 2 view geometries, ODT based on the first Born approximation was found to provide a subjectively better reconstructions than intensity based methods, however, erroneous ghost particles remain in the image. The ghost particles diminished using iterative ODT methods based on the Born expansion, but were found to be unstable unless refractive index changes were unrealistically small. Finally, the problem was formulated as an optimization problem with apriori knowledge of particle shape and, in this noiseless case, it was shown that there is sufficient information to eliminate the ghost particles. Further work is necessary to assess the practical feasibility of this approach. References [1] Coupland J. M. and Halliwell N. A., “Particle image velocimetry: three-dimensional fluid velocity measurements using holographic recording and optical correlation” Appl. Opt. 31 1004-8, 1992 [2] Barnhart D. H., Adrian R. J. and Papen G. C., “Phase conjugate holographic system for high resolution particle image velocimetry”, Appl. Opt. 33 7159–70, 1994. [3] Koek W. D., Bhattacharya N., Braat J. J. M., Ooms T. A., Westerweel J., “Influence of virtual images on the signal-to-noise ration in digital in-line particle holography” Opt. Express 13, 2578-2589, 2005 [4] Wolf E., “Three-dimensional structure determination of semi-transparent objects from holographic data” Optics Communications 1(4): 153-156. 1969. [5] Chew W. C., Wang Y. M., “Reconstruction of two-dimensional permittivity distribution using the distorted Born iterative method”, IEEE Trans. Med. Imaging, 9, 218-225,1991. [6] Rekanos I. T., Yioultsis T. V., Tsiboukis T. D. “Inverse Scattering Using the Finite-Element Method and a Nonlinear Optimization Technique” IEEE Transactions on Microwave Theory and Techniques, 47: 336-344, 1999. Bulyshev A. E., Souvorov A. E., Semenov S. Y., Svenson R. H., Nazarov A. G., Sizov Y. E., Tatsis G. P., “Three-dimensional microwave tomography. Theory and computer experiments in scalar approximation” Inverse Problems 16, 863-875, 2000. [7] Berenger J.P. “A Perfectly Matched Layer for the Absorption of Electromagnetic Waves” J. Comput. Phys. 114,185-200,1994. [8] Coupland J. M., Garner C. P., Alcock R. D. and Halliwell N. A. “Holographic Particle Image Velocimetry and its Application in Engine Development” Second International Conference on Optical and Laser Diagnostics, City University, London, UK, 12th September 2005 [9] Coupland J.M., “Holographic particle image velocimetry: signal recovery from under-sampled CCD data” Meas. Sci. Technol. 15, 4, 711-717, 2004. [10] Maas H. G.; Gruen A.; Papantoniou D. “Particle tracking velocimetry in three-dimensional flows. Part 1. Photogrammetric determination of particle coordinates” Exp Fluids, 15, 2, 133-146, 1993. [11] Elsinga G. E., Scarano F., Wieneke B., van Oudheusden B. W. “Tomographic particle image velocimetry” 6th Int. Symp. On Particle Image Velocimetry, 2005. [12] Sheng J., Malkiel E., Katz J., “Single beam two-views holographic particle image velocimetry” App. Opt. 42, 2, 2003. [13] Milkhail E. M., Bethel J. S., McGlone J. C., “Introduction to Modern Photogrammetry” John Wiley & Sons Inc. 2001 [14] Coupland J.M. and Halliwell N.A. "Holographic Displacement Measurements in Fluid and Solid Mechanics: Immunity to Aberations by Optical Correlation Processing" Proceedings of the Royal Society 453, 1053-1066, 1997