wiki:Tutorial/CorrelationImageVelocimetryOptimisation

Version 13 (modified by sommeria, 6 years ago) (diff)

--

TracNav(Tutorial/TOC)?

Tutorial / Correlation Image Velocimetry: optimisation of parameters

To improve the results from the previous tutorial, open again in the GUI series, and enter the file frame_1.png in UVMAT_DEMO01_pair/images. Select the ACTION 'civ_series' which opens the new GUI civ_input. You may import existing processing parameters by pushing the button ImportParam !ImportParam? at the top left of the GUI civ_input: open the parameter file images.civ/0_XML/frame_1.xml in the browser, or fill the GUI by hand as follows.

Time interval

The first parameter to adjust is the time interval between images, which should be sufficiently long to provide a displacement of a few pixels. The measurement precision is typically 0.2 pixel, so that a displacement of 4 pixels, as in the example, provides a relative precision of 5 %. A larger displacement would be preferable in terms of precision but may yield to poor image correlation and ’false vectors’. The choice of image pair is done in [!!ListPairCiv1].

Civ1 parameters

Once the image pair has been chosen, the next parameter is the size of the correlation box in both directions ([num_CorrBoxSize]_1] and _2), expressed in pixels. A smaller box of course improves the spatial resolution but it involves less statistics and false vectors may result from holes in the particle seeding. The particles are densely packed in this example, so we can significantly reduce the size from the default value [25,25] to [19,13] (in image pixels). The use of a elongated box along x allows to optimize the resolution in the direction y, to deal with the transverse shear.

The next parameters are ([num_!SearchBoxSize_1] and _2) and [num_SearchBoxShift_1], and _2, which determine the box in which the sub-image of the first frame is allowed to move to match the second frame. This can be adjusted from a prior estimate of the extrema of each velocity component (expressed in pixel displacement). Introduce [min, max] =[ -2; 6] for ’u’ and [-3; 3] for ’v’, and press the button [Search Range]. The optimum search ranges and shifts (for the given correlation box) are now displayed: [33 25] and [2 0] respectively.

The PIV operation is conveniently visualised by pressing the button [!TestCiv1] in the GUI civ. Then the image appears in a new GUI view_field, in which the mouse motion displays the correlation function as a color map in a secondary window. The resulting vector is shown as a line pointing to the correlation maximum. The corresponding correlation and search boxes are shown in the image.

Unselect all operations except Civ1 and validate the parameters by pressing [OK]. Then run 'civ_series' in the GUI series and look at the result by pressing [STATUS], and open the .nc result file.

To observe the influence of the search box, come back to the GUI civ_input, set [CorrBoxSize]=[19 13] and [SearchBoxSize]=[27 25] with [Shift]=0, and visualise the result with uvmat. Many black vectors (F=-2) are obtained, showing that the search domain is too small, so that the correlation maximum is constrained by the limited search interval. Using [!TestCiv1] , it can be seen that the correlation maximum is indeed at the edge of the Search box in the main flow with u$\simeq$4 (while a gap of 2 pixels is required to properly determine the maximum without edge effect).

The parameter [num_CorrSmooth] is used to fit the correlation data with a Gaussian function to obtain the maximum with sub-pixel precision. We generally keep the default value 1, while the value 2 should be more appropriate for larger particles (with wider correlation maximum).

The parameters [num_Dx] and [num_Dy] define the mesh of the measurement grid, in pixels. Reduce them to get more vectors, but keep in mind that the spatial resolution is anyway limited by the size of the correlation box, so that velocity vectors become redondant when the sub-images highly overlap those of the neighboring vector. Then the choice Dx=Dy=10, about half the correlation box, provides a good optimum.

Finally select the Mask option like in the previous tutorial.

Fix1 parameters

Select the ’FIX1’ operation, which eliminates some false vectors using several criteria. Use the default parameters.

Patch1 parameters

Select the ’PATCH1’ operation, to interpolate the vectors and calculate spatial derivatives. First choose the default parameters, press OK, run the caluclation and visualise with uvmat. We observe that a few erratic vectors have been flagged as false (painted in magenta).

Two fields can be visualised, as selected by the menu [!VelType] in the upper part of uvmat: the initial field 'civ1' and the smoothed one 'filter1', obtained by the spline interpolation/smoothing of PATCH1. Select the option 'blank' in the menu [! TransformName?] (on the left side of uvmat), to observe fields as displacement in pixel units (not physical coordinates), which is the appropriate option to analyse PIV features.

The difference between the two fields can be directly visualized by selecting 'civ1' in the menu [!VelType] and 'filter1' in the menu [! VelType_1] just below. Adjust the scale [num_!VecScale] (value 10 for instance) to better see the difference. This is rather small (0.1 px) and erratic, except in the strong shear close to the cylinder, where it reaches a value about 0.3, so the smoothing properly reduces the noise without excessive perturbation of the velocity field itself. You can also use the scalar representation, selecting the field 'U' for both 'civ1' and 'filter1'. Projection on a line (as described in tutorial 2) is also useful to get field values on a plot.

Repeat the operations by choosing the value 100 for [FieldSmooth] instead of the default value 10. Now the smoothing effect is quite clear, widening the shear region at the edge of the cylinder. 

Now come back to the default value 10, and press the button [!TestPatch?1]. This will perform patch calculations with a range of values for the smoothing parameters, and provide the rms difference between the filtered velocity field and the initial Civ1 field. This ranges from 0.12, !FieldSmooth?=1, to more than 0.2 for !FieldSmooth=100. The value 0.15 for !FieldSmooth=10 is less that the expected error on the PIV, about 0.2 pixel. 

Figure: GUI civ_input + graphe obtenu par !TestPatch.

This test also provides the proportion of excluded vectors (marqued as false) by the criterion of the excessive difference between the Civ1 value and the filtered one, which is attributed to false vectors. The threshold (expressed in pixels) is given by the box [num_!MaxDiff]. The result obtained  with the default value 1.5 is about 2 %, so that most vectors are preserved. Repeating the test  with a higher value, for instance 10, logically reduces the number of rejected vectors, but significantly increases the rms difference: the interpolation is perturbed by a few erratic vectors. 

The last parameter for Patch1 is [num_! SubDomainSize?] which corresponds to a partition in subdomains for the spline calculation, in order to avoid computer memory overflow in  the spline calculation. In this case the default choice 100 leads to a single domain.

Civ2

Select the ’CIV2’ operation to improve the correlation results, using the information on local image deformation, provided by the previous knowledge on velocity spatial derivatives (calculated in patch1). Use a finer grid dx= dy=5 than for civ1. The spatial resolution can be slightly improved by decreasing the correlation box, using for instance Bx,By=(15,11). The shift of the search range is here given at each point by the prior estimate from Civ1, so that the search range can be optimized: choose [21,17] which provides a margin of 3 pixels on each side of the correlation box. Note that ’civ2’ corresponds to a new measurement from the images, the previous civ1 and patch1 operations being used only as an initial guess for the search of optimal correlations.

FIX2 and PATCH2

Then select ’FIX2’ and ’PATCH2’ with the default parameters.

Further Civ iterations

The parameters of a CIV computation are stored in a xml file with extension ..CivDoc.xml created in the directory containing the velocity files. These parameters can retrieved, opening this xml file with the browser of the GUI civ. Then the image file itself needs to be opened (the select again the check boxes for the operations beyond civ1 hidden by default).

The result can be improved again by performing a third civ iteration, civ3. For that purpose, select only the ’civ2’, ’fix2’ and ’patch2’ operations with the same parameters as previously. The previous result is now considered as ’civ1’, so set CIV as the subdirectory in the edit window [SubDirCiv1]. Select a new subdirectory name, for instance ’CIV3’ in the edit window [SubDirCiv2]. Further iterations could be similarly performed, but the improvement becomes negligible.

The following figure shows the final vorticity field, in which the vorticity roll up in the wake of the sphere is clearly visible. A zoom near a vortex shows the vorticity superposed with velocity vectors.

Attachments (6)

Download all attachments as: .zip