Part V · Practice

15. Speckling a specimen

Chapter 1 explained why a speckle pattern is necessary at all: a featureless surface has no unique texture to lock onto. This chapter is about what makes one pattern actually work well and another one fail quietly.

What a good pattern needs

• Random, not periodic. A repeating pattern - an evenly spaced grid of dots, a woven fabric texture - gives the correlation search more than one place that looks like an equally good match. The search can lock onto the wrong repeat of the pattern with high apparent confidence, which is a far worse failure than an honest low correlation score: it reports a plausible, wrong answer instead of admitting doubt.
• High local contrast. The correlation is only as good as the intensity gradient inside each subset. A washed-out, low-contrast pattern gives the search little to grip, which is exactly what a noise floor (Chapter 9) measures directly: weak local contrast against a given camera's own noise level produces a correspondingly coarse noise floor, however good the pattern looks to the eye.
• Feature size matched to the subset, not to the whole image. What matters is not how the pattern looks across the whole specimen, but how much distinct texture falls inside one subset. A pattern with speckles far larger than the subset gives each subset too little variation to be unique; a pattern with speckles far smaller can blur together at the imaging resolution actually being used. Chapter 16 covers choosing the subset itself; the two decisions have to be made together, not in either order alone.
• Genuine full-range contrast, without saturating. A pattern that is mostly pure black and pure white loses gradient information exactly where the camera clips - a saturated pixel carries no information about how much brighter or darker the true value was, so a pattern that clips throws away contrast it could otherwise have used.

How a pattern gets onto a specimen

Some specimens already carry enough natural texture to correlate well and need nothing applied at all - worth checking, with a real photograph and a real quality estimate, before assuming a pattern is necessary. Where one is needed, common methods include spray or airbrush application of a fine random speckle, and printing a computer-generated, non-periodic pattern onto a decal or stencil and transferring or applying it to the specimen.

Checking a pattern before committing to a whole test

The cheapest check is the one done before any load is applied at all: photograph the specimen as speckled, and look at the reliability figures a correlation of that reference image against itself, or against a first small step, actually produces. A live estimate of what a chosen subset radius can resolve against the pattern actually present is worth having on screen while those settings are still being chosen, precisely because it is far cheaper to re-speckle a specimen before a test than to discover a pattern was too weak only after the test is finished and cannot be repeated.

16. Choosing subset radius and grid step

Two numbers control almost every correlation: how large a patch of pixels each measurement point looks at, and how far apart the measurement points are spaced. Neither has a single correct value; both are trade-offs that depend on the pattern and on what is being measured.

Subset radius: robustness against locality

A larger subset contains more texture, which generally makes its correlation more robust and its noise floor finer (Chapter 9) - more pixels means more gradient energy to work with, and a steadier average against sensor noise. The cost is spatial: a large subset blurs the field, because a single reported point is really an average over a fairly wide patch of the specimen. Where the deformation changes sharply over a short distance - near a hole, at a crack tip, across a narrow gauge section - too large a subset averages the interesting detail away before it is ever reported.

The panel puts a number on that cost, beside the subset radius: each displacement is an average over its subset, 2r + 1 pixels across, so 33 px at the default radius of 16. Strain is averaged further. It is fitted from the points inside the strain subregion, each of which is itself a subset's average, so it covers the span between the outermost points the fit uses plus one subset: the virtual strain gauge, 83 px at the defaults. Both lengths are the scale of the averaging, not the smallest feature that is measured at full size. A feature at that scale is measured, but smaller than it is: on a displacement wave of known size, a 33 px subset keeps 90 per cent of its amplitude only once the wavelength reaches about 130 px. Read the two lengths as the other half of the noise floor's trade: one says how finely movement is measured, the other how widely it is averaged to get there.

A smaller subset can follow sharper local detail, but has less texture to work with, is more sensitive to noise, and needs a correspondingly finer speckle pattern to still contain enough unique contrast to correlate reliably at all.

Grid step: sampling density against cost

The grid step is simply how far apart the measurement points are placed. A finer step gives a denser field and a smoother-looking result, at a roughly proportional cost in computation. It does not, on its own, improve the quality of any individual point's measurement - that is entirely the subset's job - so a very fine grid step over an unreliable subset produces a dense field of unreliable answers, not a genuinely more precise one.

Strain is the exception. Its fit uses the points that fall inside the strain subregion, so the step decides how far out the outermost of them sit, and with it the length of the virtual strain gauge. The panel states that length under the strain settings and updates it as the step, the subregion or the subset changes.

Working from the reliability figures, rather than guessing

The two settings are not really independent, and neither is right or wrong in isolation: the sensible way to choose them is to look at what the actual pattern, at the actual imaging resolution being used, can resolve at a candidate subset radius - the noise floor from Chapter 9, computed live against the settings under consideration - rather than picking a subset radius by habit or by copying a value from an unrelated test.

Seeing the numbers on the specimen

Both settings can also be drawn over the reference image, at the size they will actually be measured at, from the switches beside them on the Analysis panel. The square is the subset the correlation will solve - two radii plus one pixel across, its own centre pixel included - and the circle is the subregion the strain fit will span. Both are placed at the nearest point the run would really measure, and they follow the pointer over the image, so a candidate setting can be put over the part of the pattern that worries you and looked at.

This answers a question no number can: whether the pattern inside the square is distinct - whether there is enough unique texture in that particular patch of speckle for the correlation to have anything to lock onto. It does not replace the reliability figures above. The noise floor is the rigorous answer to how finely this subset could resolve; the box is the fast answer to whether it is looking at a pattern at all.

The points the strain fit would use are marked inside the circle, and the panel states how many there are. At a fine grid step there can be dozens of them inside a circle a few dozen pixels across, so the marks themselves appear only once the view is zoomed in far enough to tell them apart - the count beside the switch is exact either way. Near the edge of the image, and beside a hole, you will see genuinely fewer: the fit has fewer there too, which is the same lopsidedness Chapter 2 describes.

Settings you come back to

Once a subset radius and a grid step suit your specimens, they can be where every new project starts. Set them on the Analysis panel, open File > Settings, and keep the panel's settings as the defaults. The window lists every run setting with the value a new project starts at beside SurView's own, and marks the ones that differ; Restore SurView's own defaults forgets them again.

A project you open keeps the settings it was saved with, whatever the defaults are, so a saved session measures the same way on a colleague's machine as on yours. And when the window starts from saved defaults that differ from SurView's own, the log says which, so a 20 px subset chosen weeks ago never reads as the standard one.

17. Lighting and imaging

Everything in Parts I through IV assumes the two photographs differ only by the deformation being studied. Lighting, focus, and camera settings that drift between reference and target introduce differences that have nothing to do with the specimen, and DIC cannot tell those apart from real movement.

Even, diffuse, and stable

Direct, hard lighting produces specular highlights - small, very bright patches that shift position as the specimen itself moves, even though the surface underneath is speckled normally. A highlight sliding across a subset between reference and target looks, to a correlation search, like a change in the pattern itself, and can corrupt an otherwise perfectly good measurement. Even, diffuse lighting, without a strong single source, avoids this.

Lighting also has to stay stable for the whole duration of a test, not merely at the moment each photograph is taken. A light that dims, flickers, or casts a moving shadow across the specimen over the course of a long test introduces intensity changes a correlation search has no way to separate from real deformation.

Fixed focus, fixed aperture, fixed everything that can auto-adjust

Any camera setting that can change automatically between shots - autofocus, auto-exposure, auto white balance - is a setting that can quietly change the image in a way indistinguishable, to the correlation, from real deformation. All of them should be fixed manually before a test begins and left untouched for its entire duration, including through however many load steps and however much time the test takes.

The camera itself has to hold still

Everything this manual has said about a stable coordinate frame (Chapter 4) assumes the camera did not move between the reference photograph and any target. A tripod or rigid mount that cannot shift, even slightly, over the course of a test is not a minor convenience - a camera that moves introduces a rigid displacement across the entire field that is real, in the sense that it genuinely happened, and completely uninteresting, in the sense that it has nothing to do with the specimen being studied, and no way to tell the two apart after the fact.

Reading the exposure off the picture itself

All of the above is judged from the photograph, and the point panel reads it out pixel by pixel from the moment an image is loaded, before any correlation exists. Moving the pointer over the image reports what the camera actually recorded there: the value the file holds, the type it is stored in, and the range that type allows. Clicking pins the reading so it can be looked at rather than chased.

The number is the file's own, not the brightness on screen. Scientific images routinely occupy a small part of their type's range - a 16-bit image sitting in 0 to 5000 would render very nearly black - so the viewport stretches the intensity window for display. What correlation works from is the file's value, so that is what is reported, with the stretch stated beside it.