1 / 19

FLCT: A Fast, Efficient Method for Performing Local Correlation Tracking

FLCT: A Fast, Efficient Method for Performing Local Correlation Tracking. George H. Fisher Brian T. Welsch Space Sciences Laboratory, UC Berkeley. What is the basic idea behind local correlation tracking (LCT)?. V (x,y). I 1 (x,y,t 1 ). I 2 (x,y,t 1 + δ t).

fauve
Download Presentation

FLCT: A Fast, Efficient Method for Performing Local Correlation Tracking

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. FLCT: A Fast, Efficient Method for Performing Local Correlation Tracking George H. Fisher Brian T. Welsch Space Sciences Laboratory, UC Berkeley

  2. What is the basic idea behind local correlation tracking (LCT)? V(x,y) I1(x,y,t1) I2(x,y,t1+δt) Given a pair of 2D maps (“images”) of some scalar quantity, with the 2nd image taken slightly later than the 1st one, what is the 2-d flow field which, when applied to the scalar field in the first image, will most closely resemble the 2nd image? This definition of LCT is not a precise one, and in fact the problem is mathematically under-determined. In Solar Physics, the idea for LCT is generally attibuted to November & Simon (1988, ApJ 333, 427). In the engineering literature, the problem is known as the “optical flow” problem. Here, we present the technique behind the recently upgraded “FLCT” LCT method.

  3. What is the mathematical approach used by FLCT? To construct a 2D velocity field that connects two images I1(x,y) and I2(x,y) taken at two different times t1 and t2, one must start from some given location within both images, compute a velocity vector, and then repeat the calculation while varying that location over all pixel positions. Something must be done to enhance the importance of those parts of the two images that are near the selected pixel, and to de-emphasize the parts of the image that are far away from it. FLCT does this localization by multiplying each of the two images by a gaussian of width σ, centered at pixel location (xi ,yj). We denote the resulting images as “sub-images” S1 and S2. The expressions for S1 and S2 are: The quantity σ is a free parameter in FLCT, and its optimal value changes depending on the nature of the image and the size scales present in the velocity field, as illustrated later in this poster.

  4. What is the mathematical approach used by FLCT? (cont’d) The cross-correlation function of sub-image 1 with sub-image 2 is defined by We want to find, for each pair of sub-images S1 and S2 centered at position xi,yj, the shifts δx and δy that maximize C(δx,δy). The amplitude of the shifts, divided by the time δt =t2-t1 between images 1 and 2 defines the velocity determined by FLCT: vx=δx/δt, and vy=δy/δt. FLCT uses the convolution theorem to compute C(δx,δy) using Fourier transforms. If we write F(S1)= s1(kx,ky) and, F(S2) = s2(kx,ky), where F denotes Fourier transform, then the above equation can be written where F-1 denotes the inverse Fourier Transform. The name “FLCT” stems from this approach: Fourier Local Correlation Tracking.

  5. Find the peak of C(δx,δy): As a practical matter, we find the peak of f(δx,δy) ≡ |C(δx,δy)|, rather than C(δx,δy) so that the operation doesn’t involve complex arithmetic. For notational simplicity, we henceforth use x,y for δx,δy in the following discussion. We employ a curve-fitting approach to find the peak in f(x,y) that was inspired by that of Chae (2004, LCT code written in IDL). First, since the images and sub-images are computed at discrete points in space, we first identify the pixel coordinates (xm,yn) of the largest value of f, denoting the largest value of f as f(xm,yn). Note that (xm,yn) may not be equal to (xi,yj), if the location of the peak of f(x,y) has shifted by more than a pixel in x or y. To find the peak to sub-pixel resolution, we then Taylor-expand f(x,y) to 2nd order about the (xm,yn) location, denoting the expansion as ft(x,y): where the partial deriviatives are evaluated at the point (xm,yn).

  6. Find the peak of C(δx,δy) (cont’d): At the peak, we require that the x and y partial derivatives of the Taylor expanded function ft(x,y)be 0. These conditions result in a pair of linear equations which allow us to solve for the location (xmax,ymax) of the peak: To evaluate the partial derivatives, we use standard 2nd order finite difference expressions assuming a uniform grid in the sub-images: The total shift in x,y that corresponds to the peak in f is then given by δx=(xm-xi)+(xmax-xm), and δy=(yn-yj)+(ymax-yn). Equations (5) and (6) thus provide a simple and accurate method for finding the peak of the cross-correlation function to sub-pixel resolution.

  7. Summary of the FLCT computational task: • Given 2 images I1 and I2 taken at times t1 and t2=t1+δt: • For each pixel location in I1 and I2, repeat until done: • Compute sub-images S1 and S2. • Compute Fourier transforms of S1 and S2, s1 and s2. • Compute inverse Fourier transform of s1*s2. This is Ci,j. • Compute the absolute value of Ci,j, |Ci,j|. • Compute the shifts δx, δy that maximize |Ci,j|. • Compute velocities vx=δx/δt, vy=δy/δt. • Move to next pixel.

  8. Overview of FLCT code: The FLCT code, which is currently called vel_ccor, was initially written in IDL, but has since been re-written in C for portability and speed. FLCT also uses its own endian-independent binary input format for the images and for the output velocity arrays. FLCT uses the FFTW library (version 3) for computing the Fast-Fourier Transforms needed to compute the cross-correlation functions. We have written IDL procedures which write the two images vel_ccor needs into an input file, and which read the vel_ccor output file, consisting of arrays of vx,vy, and a mask array, vm. vel_ccor was designed to be easy to run from within an IDL session. To compile the FLCT code, the only external library needed is the FFTW3 library. To download a copy of the FLCT source code and compilation instructions, along with IDL i/o procedures, and pre-compiled binary executables for Linux (x86 and x86_64), Solaris, MS Windows, and OSX) go to http://solarmuri.ssl.berkeley.edu/overview/publicdownloads/software.html . Be sure to get the C version (not IDL version), and get version test_12 or later.

  9. How to run FLCT code: The FLCT code vel_ccor is designed to be run from the command-line -- either a terminal window on a unix-like machine, from the command-prompt tool in MS Windows, or from within an IDL session using either a shell escape character or a spawn command. In all cases, the syntax for vel_ccor is as follows: vel_ccor infile outfile deltat deltas sigma -t thr -h –q Required arguments: infile - contains 2 images for local correlation tracking (to create infile use the IDL procedure vcimage2out.pro) outfile - contains output vx, vy, vm (x,y velocity and mask arrays). Mask array is 0 where velocity not computed. To read outfile use the IDL procedure vcimage3in.pro) deltat - amount of time between images deltas - units of length of the side of a single pixel(note velocity is computed in units of deltas/deltat). sigma – sub-images are weighted by gaussian of width sigma.

  10. How to run FLCT code (cont’d): Optional parameters: -t thr - if average (between the two images) absolute value of the image pixel < thr, then skip calculation (if thr is between 0 & 1, thr is assumed to be in relative units of the maximum absolute pixel value. To force thr to be in absolute units, append an 'a' to it.) If calculation is skipped, then the vm mask array is set to zero at that pixel location. It is otherwise set to 1. -h - 'hires' flag - Use cubic convolution interpolation of the f(x,y) function to re-grid it to higher resolution before finding shift. Currently, this is a seldom used option of unknown usefulness which may be dropped in the future. It is retained for heritage reasons. -q - flag to suppress printing of all non-error messages Simple Example from an IDL session: IDL>f1=randomu(seed,101,101) IDL>f2=shift(f1,1,-1) IDL>vcimage2out,f1,f2,’testin.dat’ IDL>$vel_ccor testin.dat testout.dat 1. 1. 15. IDL>vcimage3in,vx,vy,vm,’testout.dat’ IDL>shade_surf,vx IDL>shade_surf,vy

  11. How quickly does FLCT run? On a 1.6 GHz Pentium M (laptop processor) running MS Windows XP, the FLCT code vel_ccor can process roughly 200 pixels per second for images with over 40,000 pixels (time per pixel is much less than this for smaller images). The code is speeded up dramatically if the –t thr option is used, in which the velocity calculation is skipped if the image values are less than the chosen threshold. There are many avenues available for further speedup of the FLCT code. Very little experimentation has been done with multiple threads thus far. An experimental version of vel_ccor written in Fortran, using MPI has been tested on our Linux cluster grizzly, and the compute speed was found to scale almost perfectly with the number of processors.

  12. What is the difference between the “new” (upgraded) FLCT code and the old one? The main difference is in how the location of the peak of the cross-correlation function is determined to sub-pixel resolution. The old version, originally described in Welsch et al. [2004, ApJ 610, 1148], used cubic convolution interpolation to compute the cross-correlation function on a much finer grid (0.1 or 0.02 pixel spacing) and then simply found the location of the largest value within the finer grid. We (and other users) found that this resulted in serious “quantization” errors when the time between images was small enough that shifts were of order 0.1-0.2 pixels or less. The new code uses equations (5) and (6) of this paper to find the location of the peak to sub-pixel resolution. We have found this to be more accurate, and more computationally efficient than the old technique. We strongly recommend that users use the new version and not any of the old versions.

  13. Some Simple Tests of FLCT Using an MDI Magnetogram and a Rotational Flow Field: Take this Quiet Sun MDI magnetogram, and rotate it by a fixed angle about the center to generate the 2nd image. The applied velocity field then corresponds to simple, solid-body rotation, with the speed increasing linearly with radius r. How well does FLCT do at recovering the applied velocity field?

  14. FLCT inverting a 1-degree rotation: Units of velocity are in pixel widths per unit time. The straight line shows the applied speed, and the scatter-plot shows the inverted speed for all inverted pixels (only points with |B| > 10G were inverted). σ=15 was assumed. The scatter is typically 0.1-0.2, and there is no significant bias in the derived speed.

  15. Inverting a 0.01 degree rotation: Again, the straight line represents the applied velocity, while the scatter plot shows the inverted velocity. Here, the scatter is significant, and the derived speeds are biased on the low side as compared to the applied speeds.

  16. Preliminary Conclusion: FLCT results in accurate inversions for pixel shifts > 0.1 On left hand side is the applied rotational speed profile; on right hand side is the derived speed profile, with pixels not included in the velocity inversion (due to weak field strength) shown as black. Saturation scales of the two images are different, but the radial speed profile is clearly visible in the image on the right hand side. This is for the 1 degree rotation case. It is not yet clear why the .01 degree inversion is so noisy. Possible sources could be effects of magnetogram noise, the method breaking down, or interpolation error when doing the small-angle rotation.

  17. How do flows derived from white light and magnetogram images differ? FLCT derived flows from AR8210 are shown in blue, for the MDI magnetogram data, and in red, for the MDI intensity data. Clearly, there are similarities (flows at lower right of image) and differences (flows at left of image). What causes these differences? There are many possible explanations, but we note that if the scalar field intensity field is changing in ways that don’t involve the velocity, one shouldn’t expect the derived flows to be the same.

  18. How does one pick the value of σ? σ=20 σ=5 The quantity σ tells FLCT how far away from a given pixel location that image information should be used to construct the cross-correlation function. Small values of σ pick out smaller scales of motion that FLCT notices; larger values will average over smaller scale velocities. Note that using σ=5 shows supergranular motions, while σ=20 does not. The choice of σ is problem-dependent, and we would characterize choosing it as an art-form, and not a science at this time.

  19. Conclusions • Upgraded FLCT code much faster and more accurate than old version • FLCT code is fast and robust, but there are many avenues for improvement • Behavior of FLCT method with σ is still incompletely understood This work was supported by NSF-ATM, by NASA Heliophysics Theory Program, the NASA LWS program, NASA Heliophysics SR&T program, and the DoD MURI grant, “Understanding Magnetic Eruptions on the Sun and their Interplanetary Consequences”.

More Related