INSTITUTE OF PHYSICS PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY Meas. Sci. Technol. 12 (2001) 1911–1921 PII: S0957-0233(01)23191-9 Local field correction PIV, implemented by means of simple algorithms, and multigrid versions J Nogueira, A Lecuona and P A Rodrı́guez Department of Mechanical Engineering, Universidad Carlos III de Madrid, c/Butarque 15, 28911-Leganés, Madrid, Spain Received 21 March 2001, in final form 11 June 2001, accepted for publication 21 August 2001 Published 9 October 2001 Online at stacks.iop.org/MST/12/1911 Abstract Local field correction particle image velocimetry (LFCPIV), which was first presented in 1997, is the only correlation PIV method able to resolve flow structures smaller than the interrogation window. It presents advantages over conventional systems and thus offers an alternative in the field of super-resolution methods. Improvements of the initial version are likely to promote its application even further. The issues defining some of these possible improvements were already indicated in the paper that originally introduced LFCPIV, but not developed. This work presents refinements and also simplifications of the technique, so that it can be applied using current algorithms of advanced correlation PIV systems. Furthermore, these refinements reduce the measurement error and enlarge the range of application of LFCPIV. In particular, the application of the system is no longer constrained to images with mean distances between particles larger than 4 pixels. Besides that, the use of interrogation windows smaller than in its previous version is evaluated. This allows multigrid LFCPIV implementations. The results show how multigrid LFCPIV can obtain better measurements than can the usual multigrid PIV, but still the refined version of the LFCPIV technique performs even better, at the expense of a larger computing time. The performance of these methods is evaluated for synthetic and real images. This includes examples in which the ability to cope with gradients in velocity, gradients in seeding density and the presence of boundaries is highlighted. Keywords: LFCPIV, multigrid PIV, super-resolution PIV, wingtip vortex 1. Introduction Basic particle image velocimetry (PIV) is able to describe two components of the velocity along a two-dimensional domain (2D 2C PIV). It has been established as an important development method for research as well as for industry. Nevertheless, in many cases there is still a gap between the information contained in the images and what is extracted from them by current PIV systems. Much of the effort put into development of the technique has recently been focused on the extraction of all the information obtainable from the images, including the development of super-resolution systems. On the 0957-0233/01/111911+11$30.00 © 2001 IOP Publishing Ltd other hand, the development of robust algorithms able to cope with especially difficult situations (i.e. large velocity gradients, seeding inhomogeneities, the presence of boundaries, etc.) has been identified as a priority in research. Generally, all these advanced algorithms are characterized as being iterative. This means that the information from initial evaluations is used to tune or adapt the system and obtain better ones. Focusing on correlation PIV, two main branches for advanced 2D algorithms can be mentioned. (i) Multigrid PIV. The iterative cycle is used to reduce the size of the interrogation windows and also to redefine Printed in the UK 1911 J Nogueira et al their location in the PIV images. When the size of the interrogation window is reduced, a fractional pixel offset of the interrogation window is essential in order to reduce peak locking. This was experimentally found by Lecordier et al (1999) and Scarano and Riethmuller (2000) and later explained theoretically by Nogueira et al (2001). This last work also reported the necessity of great care in the last steps, recommending to switch to a symmetrical direct correlation algorithm (SDC). (ii) LFCPIV, see Nogueira et al (1999). The interrogation windows are fixed in size and location, but after each iteration the image is redefined through compensation of the particle pattern deformation caused by the velocity gradients in the flow field. This implies both displacement and deformation of the image. It is performed using the displacement field from the previous evaluation. The theoretical end of this process is when both images fully coincide, thus yielding the highest correlation coefficient. It presents the ability to resolve small structures with large interrogation windows. Owing to the weighting, the compensation and the iterative procedure, the limit of this resolution is constrained only by the Nyquist criteria applied to the mean distance between particles. The convergence of the process for every wavelength is detailed in Nogueira et al (1999). Of course, the grid node spacing should also be set to satisfy the Nyquist criteria and, even in this situation, the noise present in a PIV image restricts the number of useful iterations and thus the real achievable resolution. Further insight into this issue is offered later in the paper. The system is a development of precursory techniques (Huang et al 1993, Jambunathan et al 1995) that were intended to cope with large velocity gradients. These systems lack the high spatial resolution of LFCPIV owing to an instability related to high spatial frequency. In LFCPIV a proprietary weighting function was defined in order to avoid this. The combination of these two capabilities in LFCPIV (i.e. the ability to cope with large velocity gradients and to resolve small structures in the flow) results in a very robust highresolution technique. It is relevant that, in both branches, an interpolation of the grey level image is needed, either to compensate for the deformation of the particle pattern or to produce a fractional pixel displacement. In addition to that, LFCPIV needs to include a weighting function in the correlation algorithm. In section 2 of this work it will be stated that this is also desirable in multigrid PIV, leading to the conclusion that the aforementioned weighting is worthwhile for both kinds of system. In a few words, to implement advanced 2D algorithms in correlation PIV, in addition to the usual FFT based correlation algorithm, any of these three other ones have to be used: (i) interpolation of the grey level image, (ii) application of a weighting function and (iii) using symmetrical direct correlation (SDC) or the equivalent for the last steps of multigrid PIV. Of these three, the former two will be the only ones needed for the implementation of LFCPIV described here. 1912 Consequently, this refinement with respect to the previous version (Nogueira et al 1999) avoids the use of more complicated algorithms and, thus, avoids the corresponding requirements. Owing to the relation between the present work and the one just mentioned and to allow direct comparison, the synthetic images used here will be the ones from that work if not specified otherwise. In these images, the mean distance between particle images is δ = 4.5 pixels (i.e. 4/(π δ 2 ) ∼ 0.06 particles per pixel). The mean diameter of particle images is d = 4 pixels. Their shape is Gaussian, being the diameter associated with e−2 times the maximum value. Where particles overlap, the corresponding intensities are added. 5% of particles have no second image to correlate because of their out-of-plane velocity. 2. Errors and instabilities associated with image correction in multigrid PIV The idea of relocating the information of the original image in order to optimize the recognition of the PIV correlation peak is not new. Keane and Adrian (1993) proposed the use of a discrete offset between the pair of interrogation windows. Also, the correction of the particle pattern deformation was implemented, in each interrogation window, by Huang et al (1993). Later, Jambunathan et al (1995) presented a more efficient computing method by correcting the whole image after each iteration. Both reported instabilities in their iteration and, to avoid them, were forced to use a lowpass filter and/or reduce the number of iterations to a few. It was not until the work of Nogueira (1997) and Nogueira et al (1999) that the source of the instability was identified. These works also introduced the way to avoid it without losing high-spatialfrequency information. A detailed description can be found in the mentioned works. Here, only a sketch of this relevant issue is drawn. 2.1. Sources of error and instability To start with, it is of interest to remind the reader that the measurement in a usual PIV interrogation window is related to the most frequent particle displacement, independently of its position. The contributions of the various displacements within the interrogation window have different effects depending on the ratio between the range of these displacements and the diameter of the particle images. If this ratio is smaller than unity, the peaks in the correlation domain overlap and a weighted average is obtained. For larger ratios, the overlapping of peaks is reduced, so the bias towards the most frequent displacement is increased. An extreme case would correspond to a step discontinuity in the flow field where two separated peaks could appear. As a consequence, for certain spatial wavelengths like the one represented by the sinusoidal displacement field, s, depicted in figure 1, the measured displacement at the centre of the interrogation window is opposite to the central one. In a multigrid system, this effect causes a displacement of the interrogation window in the opposite direction to that which corresponds to its centre, thus reducing the signal-tonoise ratio and possibly leading to the formation of a spurious Local field correction PIV ⇒ Measured displacement Displacement field, s. Figure 1. An example of measurement error in the interrogation window of a conventional PIV system due to high-spatial-frequency content in the displacement field. Multigrid PIV with compensarion of part. pattern and no weigh. func. (F = 16, ∆ = 4) 0.6 0.5 Multigrid PIV with compensarion of part. pattern and no weigh. func. (F = 16, ∆ = 8) 0.4 0.3 0.2 LFCPIV 0.1 0.0 1 1.2 5 9 13 17 21 25 29 33 37 41 45 Number of iterations 1.0 Figure 3. Examples of divergence for some systems with compensation for the particle pattern deformation. (F and are given in pixels.) 16 pixels window side 0.8 32 pixels window side 0.6 r Evolution of error with iterations 0.7 rms(e ) (pixels) ⇒ Instability in case of iterative compensation of particle pattern deformation 64 pixels window side 0.4 0.2 0.0 −0.2 −0.4 ∞ 64.00 32.00 21.33 16.00 12.80 10.67 9.14 8.00 λ (pixels) Figure 2. The frequency response as a function of the spatial wavelength. The zones with r < 0 indicate zones with the effect in figure 1. (Note that the scale of λ is not linear, actually it corresponds to a linear scale in frequency.) vector or outlier. To generalize the idea depicted in figure 1, figure 2 presents the frequency response, r, to 1D sinusoidal fields of various spatial wavelengths, λ, for several window sizes. Figure 2 is just an idealization, because the quantitative value of r depends on more parameters than just the frequency. However, the zones where r < 0 (which corresponds to the phenomenon depicted in figure 1) are fully correct. These zones are the source of the instability, as figure 1 indicates. If a correction of the particle pattern is attempted with this erroneous information, the deformation increases instead of decreasing. A diverging process is triggered if successive iterations are implemented. 2.2. Effects in multigrid PIV Summarizing the previous section, the error from negative frequency response leads to two effects. (i) In a conventional multigrid system it leads to a reduction of the signal-to-noise ratio instead of an increase. This boosts the possibility of obtaining outliers, specially if the interrogation window in the following iteration is smaller. (ii) In a multigrid system with correction of the particle pattern deformation (Fincham and Delerce 2000, Scarano and Riethmuller 2000), if no weighting function is used (like in the works by Huang et al (1993) and Jambunathan et al (1995)) there is an source of instability to take into account. In this section, the magnitude and effect of this error are analysed, as is the question of whether the use of an appropriate weighting function would improve the results. This analysis is presented in two parts. First, the necessary conditions for the instability to develop are studied. After that, the uncertainty due to this error, even when the instability is not allowed to evolve, is analysed for various wavelengths. The basic condition for the unstable growth is that the grid-sampled displacement field contains the required (r < 0) frequencies. The low resolution of CCD cameras gives this condition except for very simple flows. Besides that, image noise and image discretization introduce frequencies that could be troublesome. Last but not least, particles act as samplers of the flow field, introducing aliasing if they are too far apart. Nevertheless, following figure 2 and taking into account the Nyquist criteria, the grid-sampled displacement field will contain unstable frequencies only if the grid node spacing, , is smaller than half the length of the side of the square interrogation window, F . To illustrate this behaviour, a pair of synthetic images of a field with uniform displacement of 3.6 pixels, d = 2 pixels and δ = 1.4 pixels was analysed. The results are plotted in figure 3. These synthetic images contain absolutely no frequencies related to the instability except for the noise due to discretization of grey levels. It can be seen that a multigrid system with compensation of the particle pattern deformation, iterating with F = 16 pixels and = 8 pixels, does not exhibit divergence after 45 iterations. The same system with = 4 pixels exhibits it clearly. The performance of the LFCPIV described in section 3 is also depicted for comparison. Even when this instability is not allowed to grow (large grid spacing or few iterations), the error from r < 0 affects the accuracy. To analyse this accuracy, let us define the usual parameters for a multigrid PIV implementation. A representative example is a five-iteration algorithm. The size of the interrogation windows in this test will successively have the following values: 64 pixels for the first iteration, 32 pixels for the second one and 16 pixels for the last three. The space between interrogation grid nodes will be successively 16 pixels, 8 pixels and three times 4 pixels. Correction of the particle pattern deformation will be applied after each iteration. Later on, a discussion of the variations of these parameters is offered. 1913 J Nogueira et al The concepts introduced in figures 1 and 2 help us to understand the output after each iteration. The detailed analysis of the proposed multigrid system measuring a sinusoidal 1D field of wavelength λ = 21.3 pixels, on the synthetic images defined in the introduction, is as follows. 0.50 0.25 υ2 0.00 -0.25 -0.50 -0.25 0.00 ξ/F 0.25 -0.50 0.50 η/F Figure 4. The weighting function designed to avoid negative frequency responses. For details see Nogueira et al (1999). Displacement field: s = 2sin(2πx /λ x ) (pixels) 1.2 Multigrid PIV+compens. (F :64,32,3x16; ∆ :16,8,3x4) 1 rms(e )/rms(s ) (i) For the first iteration (F = 64 pixels), figure 2 shows that r is slightly negative. The result is that this first iteration gives in some places a displacement opposite to the one being measured. If we analyse the root mean square (RMS) value of the error, e, in the measurement, normalized with respect to the RMS of the signal to be measured, the value obtained is RMS(e)/RMS(s) = 1.07. This value is consistent with a measured field slightly opposite to the real one. (ii) For the second iteration (F = 32 pixels), figure 2 shows a highly negative value for r. The result in error terms is RMS(e)/RMS(s) = 1.34. This is consistent with a measurement with opposite sign to the original displacement field. (iii) In the three following iterations (F = 16 pixels) the value of r is positive. The result is a reduction of the error in each step, finishing with RMS(e)/RMS(s) = 1.19. Further iterations would lead to worse measurements due to accumulation of errors from other sources. 8-9 7-8 6-7 5-6 4-5 3-4 2-3 1-2 0-1 Multigrid PIV+compens.+weight in first 2 steps (F :64,32,3x16; ∆ :16,8,3x4) (Multigrid LFCPIV in the first two steps) 0.8 0.6 0.4 0.2 2.3. Solutions for multigrid PIV The conclusion from the test in the previous section is that the effect of the error mentioned can be relevant. One solution for this would be to use only 16 pixel by 16 pixel interrogation windows. This would avoid the build-up of error down to λ = 16 pixels. The performance of such a system has already been studied with different implementations by Jambunathan et al (1995) and Scarano and Riethmuller (2000). But the use of just small windows for every step in a system reduces considerably the robustness against outliers as well as the dynamic range of measurable displacements. Another solution, recommended here (and already proposed by the authors in Rodrı́guez et al (2000)), is to use an appropriate weighting function to avoid the source of error depicted in figures 1 and 2. The weighting function designed for LFCPIV is able to perform this task. This weighting function is presented below and depicted in figure 4: 2 2 ξ ξ η η υ 2 (ξ, η) = 9 4 − 4 + 1 4 − 4 + 1 (1) F F F F where ξ and η are coordinates with their origins at the centre of the interrogation window and F is the length of its side. Unfortunately, the application of a weighting function induces some erroneous slippage into the measurement (Nogueira et al 1999). This slippage is larger with smaller windows, making its use for windows with sides smaller than 32 pixels inadvisable. Consequently, for the five-iteration system analysed in this section, the weighting function is used only for the first two iterations. Another option, after using this weighting function, υ, in the first two steps, is to implement also in the last three steps a 32 pixel by 32 pixel window with the weighting function, instead of 16 by 16 without it. This procedure can be defined as multigrid LFCPIV. 1914 0 Multigrid LFCPIV (F :64,4x32; ∆ :16,8,3x4) 20 25 30 35 40 45 50 λ x (pixels); (δ =4.5 pixels). 55 60 Figure 5. The performance of the five-iteration multigrid systems described in the text, together with those of the methods from Jambunathan et al (1995) (thin full line) and Scarano and Riethmuller (2000) (thin broken line). The results for all these options are compared in figure 5. In figure 5, the two thin lines plotted as references correspond to the two mentioned systems that only use 16 pixel interrogation windows. The oscillation around the reference lines of the multigrid system without a weighting function is caused by the effect depicted in figure 1 when one is operating with the two different window sizes. Details on the images used for this evaluation can be found in Nogueira et al (1999), except for the one from Scarano and Riethmuller (2000), which was directly taken from that reference. It must be accepted that the parameters defined here for multigrid PIV are arbitrary, but the benefits of the application of the weighting function for steps with windows larger than 30 pixels are obvious. It should be remarked that most of these benefits remain even when there is no compensation for the particle pattern deformation. Another useful conclusion is that better results can be obtained with weighted 32 pixel windows rather than unweighted 16 pixel windows. It should be remarked that the application of windows smaller than 16 pixels on real images is still a subject under development. Many actual applications of advanced multigrid systems do not use smaller windows. An example is Scarano and Riethmuller (2000). In that paper devoted to multigrid PIV, the processing of real images is performed with windows of 32 and 16 pixels on a side. Local field correction PIV Nevertheless, there is ongoing research on systems especially designed to deal with smaller windows, Fincham and Delerce (1997) show that good measurements can be obtained with small windows, but the requirement for a large particle diameter arises (the optimum diameter is around 6 pixels). Nogueira et al (2001) show a promising way to deal with small windows without this requirement, by the use of SDC algorithms. Besides the need for appropriate algorithms, the robustness of the system decreases strongly with decreasing size of interrogation window. These facts encourage the full development of a system with resolution smaller than the size of the interrogation window, thus not requiring size reduction. 3. A refinement of the original LFCPIV method Up to now, the only method able to obtain resolutions smaller than the interrogation window is LFCPIV. In this section, a refinement of the previous version is described. 3.1. Initial considerations The starting point of any cross-correlation PIV system is a pair of images of the particle pattern, a and b, separated by a known time interval. Cross correlation of the corresponding interrogation windows, at the measurement points, approximates the displacement field of the particles on going from one image to the other. The subsequent action of any iterative system is to use the information of this measurement to adapt the system in such a way as to obtain better ones. In particular, in a system with compensation for the particle pattern deformation, a third image b∗ is obtained by deforming and shifting b with the information of the approximate displacement field. This reduces the relative distortion, thus increasing the signal-tonoise ratio. Consequently, further analysis by cross correlating a and b∗ yields information to reduce the error of the approximated displacement field. This allows corrections that provide successive images b∗ from b to feed the iterative cycle. Details on the image processing can be found in Jambunathan et al (1995) and Nogueira et al (1999). To be able to iterate without instability due to the effect depicted in figure 1, the weighting function depicted in figure 4 has to be used (as noted in the previous section). Unfortunately this introduces a small error that accumulates through the iterations and starts to be significant after about 15–20 iterations. To avoid this error, a complex algorithm was designed at the birth of LFCPIV. The price to pay was a reduction of the field of application of the system to images with δ > 4 pixels. Many PIV images are shot in industrial wind tunnels, with a heavily seeded flow and with a strong need for high resolution. These images present small distances between particles and, actually, this means that one makes better use of the limited CCD sensor resolution. This leads to the search for new solutions in order to deal with the error introduced by the weighting function. Here a new one without reduction of the field of application is proposed. Furthermore, the results obtained with this refinement reduce the uncertainty of the measurement, particularly at high spatial frequencies. One last consideration is that, to obtain a second-order approximation of the velocity field, the distortion of the particle pattern can be compensated in both images a and b, giving a ∗ and b∗ . This way, only half of the deformation has to be compensated in each image, resulting in a system with an increased ability to cope with large velocity gradients. 3.2. A new approach to dealing with the error introduced by the weighting function Observation of the measurement slippage due to the weighting function leads to the conclusion that it is generally small, due to statistical cancellation in the interrogation window. This means that the locations affected by significant slippage are scarce. Consequently, a solution is to freeze the measurement for the few nodes with significant slippage in the ongoing iteration. This way the evolution of these nodes to give a worse measurement instead of a better one is avoided. This is the objective of the system here proposed. The detection of significant slippage is performed in an approximate way, but the results obtained demonstrate that it is a significant improvement. The nodes frozen in a certain iteration are those with slippage large enough to be detected. The detection is based on the increase or decrease of the local correlation coefficient after each compensation of the particle pattern deformation. Several sources of noise mask the detection of slippage. The main one is the influence of neighbouring nodes on the change of the local correlation coefficient. A specific procedure was designed to deal with this issue, which is described in detail in what follows. 3.3. The refined system A detailed description of the refined system is specified through the following steps. (i) Calculation of local coefficients. The value of the local correlation coefficient in a window with F = 2 is calculated for each grid point. This value will be used later to search for slippage in the displacement values. This window corresponds to the region of influence of each grid node. (ii) Initial processing of the images. This step is carried out as in usual cross-correlation PIV. The image is divided into windows (usually larger than the ones in the previous step) and these are cross-correlated to find the displacement peaks. The only difference is that the weighting window depicted in figure 4 is used. The resulting expression for the cross-correlation coefficients, Clm , is specified below: F /2 Clm = υ(ξ, η)f (ξ, η)υ(ξ, η)g(ξ + l, η + m) × ξ,η=−F /2 F /2 υ 2 (ξ, η)f 2 (ξ, η) ξ,η=−F /2 × F /2 −1/2 υ 2 (ξ, η)g 2 (ξ + l, η + m) (2) ξ,η=−F /2 where f (ξ, η) and g(ξ, η) are the grey-level maps of the interrogation windows belonging respectively to the first and second images to be correlated. υ(ξ, η) is the square root of the values depicted in figure 4. 1915 J Nogueira et al The evolution of the error in these iterations is qualitatively similar to that of its former version (decreasing fast at the beginning to increase slowly after a minimum). To decide when to stop the iterations in the case F = 64, an improved method has been implemented through the following two modes of operation. (i) Mode 1 corresponds to normal iterative PIV operation. All the nodes are considered for evolution. When the number of worsening nodes is more than half the number of improving ones, the system changes to mode 2. The rationale behind this is the need to detect that a large enough number of nodes have already converged. (ii) In mode 2 only the nodes with local correlation coefficients below the average, at the time of mode change, are considered for evolution. When the number of worsening nodes is more than the number of improving ones the system changes to mode 1. This mode complements mode 1 insofar as it freezes the nodes that have already converged, allowing the more slowly evolving ones to proceed. If no further iteration is made in mode 1 or 2, the system stops. In the case F = 32, the slippage introduced by the weighting function makes it advisable to take into account more conservative considerations. A reasonable option is to limit the number of iterations to a few (for example ten). In both cases (F = 64 and F = 32) the differences from the minimum value of RMS(e)/RMS(s) were smaller than 0.05. 1916 Displacement field: s = 2sin(2πx /λ x ) (pixels) 0.7 Former LFC-PIV (F = 64 pix.) 0.6 rms(e )/rms(s ) (iii) Compensation for the particle pattern deformation. Correcting the particle pattern deformation gives images a ∗ and b∗ . In the system presented here, the interpolation applied to obtain the grey levels is biparabolic. (iv) Recalculation of local coefficients. The local correlation coefficients are obtained by windowing images a ∗ and b∗ . These values are compared with the ones obtained in step (i). A lower value may mean detection of significant slippage or simply influence by neighbouring slips. In consequence, of the coefficients that worsen, only those surrounded by at least another five that also worsen, among the eight neighbours, are considered to contain significant slippage. The evolution of the nodes with significant slippage is avoided during this cycle. (v) Validation and interpolation of displacements to avoid intermediate false measurements. This step avoids obvious outliers in intermediate cycles. Any proven validation and interpolation algorithm would be valid. In particular, the ones applied here are those from Nogueira et al (1997). (vi) Compensation for the particle pattern deformation. With the approximation to the particle pattern displacement so far obtained, a new pair of images, a ∗ and b∗ , is obtained again from a and b. The local coefficients relating a ∗ and b∗ are stored like in step (i). (vii) Further processing on the images a ∗ and b∗ . These two images are fed into step (ii). The measurement obtained is a correction to the previously estimated displacement field. By adding this correction, a new approximation to the particle pattern deformation is obtained. This is supplied to step (iii), defining the iterative loop. 0.5 Refined LFC-PIV (F = 32 pix.) 0.4 Refined LFCPIV (F = 64 pix.) 0.3 0.2 0.1 0 20 25 30 35 40 45 50 λ x (pixels); (δ =4.5 pixels). 55 60 Figure 6. Performances of LFCPIV systems for the displacement fields indicated. Recent improvements in the mode of operation have shown lower error figures (Rodrı́guez et al 2001). The resulting method described so far requires only two types of algorithms: interpolation of grey levels and correlation calculation. Use of algorithms on validation and interpolation of vectors is a valuable option and almost essential for intermediate steps, if the noise or the velocity gradients are large. A detail that further increases the accuracy of the system and that will be applied in this paper is to symmetrize the correlation algorithm. An alternative formulation to expression (2) can be obtained by swapping f and g. The resulting measurement of displacement can be averaged with the one from expression (2). The performance gives an error reduction of ∼3%. The price to pay is that the computing time must be doubled. It is also of interest to note that, when the displacements to be measured are smaller than 0.5 pixels (typically after the third iteration or so), only the central correlation coefficient plus its four closest neighbours have to be calculated. For this case a direct correlation is more efficient than a FFT algorithm. Nevertheless, usual application of the system for high resolution requires more than 100 iterations with of the order of four pixels, making this a considerable computational load. It requires, like its previous version, a time of the order of 10 + (n − 5)/5 times what is required for a conventional system (n > 5 being the number of iterations). 4. Evaluation of synthetic images The synthetic images used in section 2 were also used here to allow direct comparison with the results in that section and the results from Nogueira et al (1999). In this case the performance of the system described in section 3 is plotted, in figure 6, for F = 32 pixels and F = 64 pixels. Also the performance of the previous version of LFCPIV is plotted. = 4 pixels was used in all processing. It can be observed that the case of F = 64 pixels always performs better. Images with the same attributes but d = 2 pixels and δ = 1.41 pixels (0.64 particles per pixel) were processed with even better results. This observation emphasizes that the restriction of application to images with δ > 4 pixels has been eliminated. Up to this point, the system has been detailed and initially tested. However, the data shown so far are not enough for a clear picture of the system’s behaviour. More information is Local field correction PIV AVERAGED DISPL. PROFILE < 0.5F Displacement field ⇒ 4 Measured displacement Figure 7. An example of measurement error in the interrogation window of a conventional PIV system due to non-periodic small features in the displacement field. Displacement (pixels) F 1 0 2 4 6 8 10 12 14 16 Position (pixels) rms AROUND REAL VALUE 4 rms around real value (pixels) Some tests were carried out to show that the effectiveness of the LFCPIV system is not due to these issues. The conclusions are detailed below. The presence of a weighting function reduces the effective size of the window. Nevertheless, in all the tests carried out, all the weighed windows behaved more robustly than did windows half the size with no weighting applied. This means that the number of outliers was smaller. Therefore, the ratio (effective size)/(real size) can be considered larger than a half. In comparison with this reduction of effective size, the increase in spatial resolution is much larger. The weighted windows in a LFCPIV system are able to resolve features in the flow with spatial wavelengths smaller than F /4. Consequently, the increase of resolution cannot be attributed to the reduction of the effective size. Moreover, figure 6 shows that a larger window performs better than a smaller one in terms of accuracy. Although some other effects have to be taken into account, this indicates that the benefits in terms of resolution are not coupled to a reduction of effective size. Concerning the issue of periodicity in the test flow fields, in order to check whether the periodicity of the field is helping the algorithm to work, several tests were performed using nonperiodic fields. All of them gave similar results to those using periodic fields, showing that the capability of the LFCPIV system to achieve high resolution is not a consequence of the existence of periodicity in the flow. Before commenting on some of these tests, it is interesting to point out some considerations about non-periodic fields. The foreseeable detectability threshold of a perturbation in a homogeneous field is depicted in figure 7, for a conventional PIV system. If the area covered by the perturbation is smaller than 0.5F 2 , its peak in the correlation domain would be smaller than that belonging to the unperturbed field. Consequently, if the two peaks do not overlap, the perturbation would not be detected. If they overlap, the one associated with the unperturbed field would be dominant. Instead of this threshold, a LFCPIV system has a value of ∼0.21F 2 , thanks to the weighting function. The same reasoning can be applied for a 2D discontinuity. In addition to this, an iterative system like LFCPIV would not lose track of the feature in cases in which the peak of the unperturbed flow in the correlation domain overlaps with the peak corresponding to the 2 0 needed in order to confirm that the increase of accuracy, with respect to that of a conventional PIV system, should not be erroneously attributed to the following two features: (i) the reduction of the effective size of the window and (ii) the periodicity in the test displacement fields. 3 3 2 1 0 0 2 4 6 8 10 12 14 16 Position (pixels) Figure 8. Performances of two systems for measurements of small non-periodic features in the displacement field. Thick line, field to be measured; circles, LFCPIV measurements (F = 64 pixels); and squares, multigrid PIV (F = 16 pixels). The RMS error of the various measurements around the real value is also depicted, to give a complete picture. perturbation. Examples of this circumstance appear when the step depicted in figure 7 has a finite slope or when the increment in displacement is smaller than d. In these cases, the iterations would lead the system to the right solution. To illustrate this effect, a pair of synthetic images of a zero displacement field crossed by a perturbation with a parabolic profile (laminar Poiseuille 2D flow) was generated and analysed. The width of this perturbation is eight pixels and the maximum value of the displacement is four pixels. The images were generated with δ = 1.4 pixels and d = 2 pixels. These images were analysed using the LFCPIV system with F = 64 pixels. This means that the width of the perturbation is much smaller than the foreseeable threshold of detectability defined above. However, the finite slope introduces information in the larger peak that slowly guides the iteration to the right solution. A multigrid system with F = 16 pixels was implemented for comparison. This system is equivalent to the ones in the caption of figure 5. The results are depicted in figure 8. Another example of a non-periodic field, that highlights other features, such as the presence of boundaries and seeding density inhomogeneity, follows. Analogously to the previous case, it concerns a parabolic profile with a maximum displacement of four pixels and the same values for δ and d. In this case the width of the profile is 16 pixels and there is no seeding outside it. A couple of 64 pixel by 64 pixel images to be correlated are depicted in figure 9. 1917 J Nogueira et al Figure 9. An example of 64 by 64 interrogation windows, a and b, of a displacement field that corresponds to a horizontal Poiseuille flow. The width of the flow is 16 pixels. AVERAGED DISPL. PROFILE Displacement (pixels) 4 Figure 11. An image of the central zone of a wingtip vortex analysed in section 5 (provided by the DLR). 3 5. Application to real images 2 1 0 0 2 4 6 8 10 12 14 16 Position (pixels) rms AROUND REAL VALUE rms around real value (pixels) 3 2 1 0 0 2 4 6 8 10 12 14 16 Position (pixels) Figure 10. Performances of two systems for measurements of small non-periodic features and boundaries in the displacement field. Thick line; field to be measured; Circles; LFCPIV measurements (F = 64 pixels); and squares; multigrid PIV (F = 16 pixels). It can be seen that no reflection of light was included at the boundaries. Nonetheless, the case is tough due to the tendency of conventional systems to delocalize the information from particles inside the interrogation window. The performances of the same two systems are depicted in figure 10. One last observation is that, in the LFCPIV system presented here, even in the case of losing the information on a feature that could be resolved with the grid spacing, the local coefficient defined in the previous section would inform us about it. With this information, the time between pulses can be reduced and the feature would be more easily tracked. Therefore the LFCPIV system has a high robustness besides a high accuracy, beyond those of other current high-resolution techniques. 1918 As an example of application of this version of LFCPIV to real images, a typical example from large-scale aerodynamic facilities has been chosen. The particular case corresponds to the core of a wingtip vortex generated in the DNW wind tunnel with a fixed model of half an airplane wing, corresponding to a wingspan of 3.5 m, in a test section of 6 m × 8 m. The PIV image to be analysed was obtained with the DLR’s standard PIV system comprising Nd:YAG lasers (2 × 600 mJ), Laskin nozzle-type aerosol generators and a high-sensitivity cooled ‘cross-correlation’ camera (1280 × 1024 pixels). The time between images was 2 × 10−5 s. This image, which is one example out of several thousand PIV recordings, has been obtained by the DLR within the EC-funded EUROWAKE project dedicated to the investigation of wake vortices. In the frame of our work it is studied rather from the point of view of PIV evaluation methods than of drawing conclusions with respect to the behaviour of wake vortices. Further detailed information about this experiment can be found in Kompenhans et al (1999). For clarity, only the analysis of the centre of the image will be presented here (216 × 208 pixels). One image of this zone is depicted in figure 11. As has been noted, this image presents a number of characteristics associated with large industrial facilities that made it interesting to test this version of LFCPIV on it. In some zones, δ is smaller than two pixels. This allows us to show that the new version of LFCPIV has no intrinsic limitations regarding the distance between particle images. Simultaneously, it presents high gradients in seeding density (from 0.6 to 0.1 particles per pixel) and velocity (up to 3.2 × 104 s−1 ). This allows us to show the ability of the system to deal with such difficulties. The measurement is performed with F = 64 and = 8 pixels. Processing with conventional PIV systems gives around 60–100 outliers near the core. The measurement performed with LFCPIV, depicted in figure 12, gives no final outliers. On the other hand, three of the central measurements are questionable due to its low local correlation coefficient (<0.5). In the zone where these measurements are located, the Local field correction PIV : 6 pixels. (a) (b) Figure 14. SNRs for data in figure 12 (>2, white; 2 to 1, grey; <1, black). (a) Only a discrete offset of interrogation windows. (b) Compensation for the particle pattern deformation. Figure 12. A measurement carried out with LFCPIV on the image in figure 11 and its pair ( = 8 pixels). Vorticity (1/s) 21000-22500 19500-21000 18000-19500 16500-18000 15000-16500 13500-15000 12000-13500 10500-12000 9000-10500 7500-9000 6000-7500 4500-6000 3000-4500 1500-3000 0-1500 -1500-0 25 22 19 16 13 10 1 7 6 11 column 4 16 21 row use interpolation of the image grey levels. It is true that interpolation means a degradation of the quality of the image. Sometimes, this fact is a source of doubts about the convenience of its application. The image analysed in this section suits as an example to show that the benefits of the compensation for the particle pattern deformation can be much more significant than the possible degradation. To illustrate this, the signal-to-noise ratio (SNR) of local 32 by 32 windows will be studied. The SNR is usually plotted as the ratio (signal peak)/(highest noise peak) in the correlation domain. Here, the differences in seeding density among parts of the image would fool this indicator because of the differences in the mean value over which the peaks arise. To allow a more uniform comparison along the image, an alternative ratio is proposed: signal peak − average value . highest noise peak − average value (3) 1 26 Figure 13. A vorticity plot of the vectors depicted in figure 12. The centre peak has been cut (white feature in the centre) due to the questionable nature of the results and to clarify the presentation. velocity gradients along streamlines are too large to consider the time between images appropriate (∼2 × 104 s−1 ). This means that the system is failing only where the measurement is not valid for a reason external to the algorithms, which is the time between pulses. This difficulty may also be combined with a low seeding density (owing to centrifugal migration) and the possible existence of an out-of-plane displacement. It should also be remarked that the displacement field is not as smooth as it seems visually in figure 12. Figure 13 shows that there is a ring of vorticity of characteristic width of the order of 32 pixels that is clearly described. This again highlights the capability of the system to obtain measurements of features smaller than the interrogation window, even under adverse conditions. Whether the vorticity peaks in this ring are caused by peak locking is not clear, but revealing them is a consequence of the high resolution capability of the method in this case. The vorticity algorithm applied here is a usual circulation filter of size 2. This allows comparison with standard methods. Nevertheless, more advanced algorithms can be found in Lecuona et al (1998). One last issue about real images can be remarked upon. As was stated is section 1, all the advanced 2D PIV algorithms In this expression the average value is obtained in the correlation domain using all the values smaller than the smallest of the two peaks defined. Using the displacement field presented in figure 12, this signal-to-noise parameter is plotted in figure 14 for each grid node, allowing comparison between two cases: (i) a discrete offset of windows and (ii) compensation for the particle pattern deformation. In figure 14, SNR < 1 means that the noise peak is higher than the signal peak. This could just indicate a different value from what has been considered as the true displacement (based on the solution from LFCPIV). About this, it is important to mention that the highest peak belonged to a clear outlier in more than 50% of cases, reinforcing the meaning of SNR < 1. A similar analysis was performed for 16 by 16 and 64 by 64 interrogation windows with even more evident differences between the two methods. To further illustrate this effect, figure 15 shows a set of interrogation windows and the corresponding correlation domains. These interrogation windows come from the grid node of coordinates (9, 9) with its origin in the lower left-hand corner of figure 12. 6. Conclusions A refined LFCPIV method that uses only current algorithms of any advanced 2D PIV system has been developed and its 1919 J Nogueira et al (I) image a image b (II) image a* image b* Figure 15. The effect of compensation. (I) From left to right: interrogation window from image a, interrogation window from image b and corresponding correlation. Only a discrete offset of interrogation windows has been used. (II) The same as before but with compensation for the particle pattern deformation. metrological quality tested. The new system presents some advantages over the previous version. There is no restriction on the mean distance between particles. Smaller interrogation windows can be used. It provides a lesser measurement uncertainty than that of its predecessor. The result is a robust high-resolution system able to cope with large seeding densities and velocity gradients. It has been shown that the provision of high resolution is not coupled either to a decrease in the effective size of the interrogation window or to the periodicity of the displacement fields on which the previous version was tested. A multigrid version that yields results better than standard multigrid systems down to 16 pixel by 16 pixel interrogation windows has also been presented. The price to pay is an increase in computing time by a factor of six. Recent systems based on windows smaller than 16 pixels by 16 pixels, like the SDC algorithms mentioned, probably will give good results in the future. They can always be implemented after a LFCPIV system instead of a multigrid system down to 16 pixels by 16 pixels. This version would obtain benefits from the robustness against outliers of LFCPIV. Although both SDCPIV and LFCPIV algorithms have already presented clear improvements, ways for further refinement of their concepts still remain to be investigated. Interpolation of the PIV image is needed both for multigrid systems and for LFCPIV. It has been shown that, when high velocity gradients are present, the SNR can be significantly increased thanks to the interpolation. It should be noted that the interpolation in each iteration is always based on the original image and there is no interpolation of images that have already been interpolated, for that would reduce substantially the quality of the image. 1920 Acknowledgments This work has partially been funded by the Spanish Research Agency grant DGICYT TAP96-1808-CE, PB95-0150-CO202 and under the EUROPIV 2 project (a joint programme to improve PIV performance for industry and research), which is a collaboration among LML URA CNRS 1441, Dassault Aviation, DASA, ITAP, CIRA, DLR, ISL, NLR, ONERA, DNW and the universities of Delft, Madrid (Carlos III), Oldenburg, Rome, Rouen (CORIA URA CNRS 230), St Etienne (TSI URA CNRS 842) and Zaragoza. The project is managed by LML URA CNRS 1441 and is funded by the CEC under the IMT initiative (contract GRD1-1999-10835). References Fincham A M and Delerce G 2000 Advanced optimisation of correlation imaging velocimetry algorithms Exp. Fluids. 29 S13–22 Huang H T, Fiedler H E and Wang J J 1993 Limitation and improvement of PIV (part II: particle image distortion, a novel technique) Exp. Fluids. 15 263–73 Jambunathan K, Ju X Y, Dobbins B N and Ashforth-Frost S 1995 An improved cross correlation technique for particle image velocimetry Meas. Sci. Technol. 6 507–14 Keane R D and Adrian R J 1993 Theory of cross-correlation of PIV images. Flow Visualization and Image Analysis ed F T M Nieuwstadt (Dordrecht: Kluwer) pp 1–25 Kompenhans J, Dieterle L, Vollemers H, Stuff R, Schneider G, Dewhirst T, Raffel M, Kähler C, Monnier J C and Pengel K 1999 Aircraft wake vortex investigations by means of particle image velocimetry: measurement technique and analysis methods 3rd Int. Workshop on PIV’99, University of California, Santa Barbara Local field correction PIV Lecordier B, Lecordier J C and Trinité M 1999 Iterative sub-pixel algorithm for the cross-correlation PIV measurements 3rd Int. Workshop on PIV ’99. University of California, Santa Barbara Lecuona A, Nogueira J and Rodrı́guez P 1998 Flowfield vorticity calculation using PIV data J. Visualization 1 183–93 Nogueira J 1997 Contribuciones a la técnica de velocimetrı́a por imagen de partı́culas (PIV) PhD Thesis ETSI Aeronáuticos, Universidad Politecnica de Madrid Nogueira J, Lecuona A and Rodrı́guez P A 1997 Data validation, false vectors correction and derived magnitudes calculation on PIV data Meas. Sci. Technol. 8 1493–501 ——1999 Local field correction PIV: on the increase of accuracy of digital PIV systems Exp. Fluids 27 107–16 ——2001 Identification of a new source of peak-locking, analysis and its removal in conventional and super-resolution PIV techniques Exp. Fluids 30 309–16 Rodrı́guez P A, Nogueira J and Lecuona A 2000 On the design rules for PIV: limitations on the size of interrogation windows in PIV correlation systems, need for symmetrical direct correlation PIV Proc. 3rd Workshop on Particle Image Velocimetry of PIVNET. Lisbon ——2001 Modification of the Local Field Correction PIV technique to allow its implementation by means of simple algorithms Proc. 4th Int. Workshop on PIV’01, Göttingen Scarano F and Riethmuller M L 2000 Advances in iterative multigrid PIV image processing Exp. Fluids 29 S51–60 1921