Lighting for Metal Surface Scratches: Dark Field Without Guesswork

Almost every failed scratch inspection starts the same way: a bright ring light is installed, the image looks clean and evenly lit, and the scratches are still invisible. The light is not too weak. The geometry is wrong, and no amount of current will fix it.

This note is about the geometry — what dark field actually does on metal, how to set the angle instead of guessing it, and where the method stops working. It does not quote detection rates, because those belong to a specific part, a specific scratch population and a specific acceptance criterion, and none of them travel well between projects.

Why a clean, evenly lit image hides scratches

On a machined or polished surface, a scratch is a change of slope, not a change of colour. Under high-angle illumination the flat surface and the scratch flanks both send light toward the lens, so both image bright and the contrast collapses. The better the illumination — uniform, diffuse, glare-free — the more completely the scratch disappears.

Dark field inverts the geometry. The source is placed outside the specular wedge that the camera images (the “illumination W” in the standard treatment of front lighting). A flat surface reflects the grazing light away from the lens and images dark; the scratch flanks, tilted relative to the surface, scatter light back into the lens and image bright. Contrast comes from slope, so it survives on parts where brightness-based contrast has nothing to work with.

The practical test for whether a surface is a dark-field candidate is simple: if you can see a reflected lamp or window on it, dark field will work. If the surface looks uniformly matt at every angle, the same scattering that makes dark field work is already happening everywhere, and the background will rise with the defect.

Angle: choose it from the shadow budget, not from a catalogue

Low-angle illumination runs from roughly 5° to 30° above the surface. Lower angles reject flat-surface reflection more completely, which raises defect contrast; but a raised feature of height h casts a shadow of about h / tanθ, and that shadow can hide the defect behind it. At 10°, a 0.5 mm burr throws a shadow around 2.8 mm long — on a dense part that is several features wide.

So the angle is not a free parameter to minimise. It is set by two competing requirements: steep enough that shadows do not swallow neighbouring regions, shallow enough that the flat surface stays dark. In practice, start at 20–30° for parts with tall features and 10–15° for flat surfaces with fine scratches, then verify by rotating a reference scratch through four orientations.

Azimuth decides whether a scratch is visible at all

A scratch is a groove with two flanks. Light arriving perpendicular to the groove scatters strongly from both flanks; light arriving parallel to it produces mostly shadow along the groove and far less signal. A single bar light therefore has a systematic blind spot at 90° to its preferred orientation, and that blind spot is invisible in the image — the operator simply sees “no defects” on that half of the part.

The fix is not more power but more directions: opposed bars at 120–180° included angle, or a four-quadrant low-angle ring. Where a single direction is unavoidable, rotate the part or the light between captures and fuse the images. It is worth building the reference sample with scratches in at least four orientations, because a validation set with one scratch direction will pass a setup that misses half the real population.

Bars, rings and the roughness limit

Bar lights are the flexible option: one bar for a known orientation, two opposed for mixed orientation, four for an omnidirectional check. Low-angle rings are more compact and symmetric, which suits cylindrical parts where any azimuth preference will rotate with the workpiece. Both work on the same principle; the choice is mechanical.

The limit is surface roughness. Dark field depends on the flat surface staying dark. As Ra rises, more of the surface scatters at angles the lens can see, the background lifts, and contrast falls. Shot-blasted, sand-cast and heavily brushed surfaces are usually poor candidates; lapped, ground, turned and polished surfaces are good ones. Textured surfaces also produce a speckled background that defeats threshold-based segmentation — in those cases high-angle diffuse illumination with texture-suppression filtering is often more stable, even though it is optically less elegant.

What to measure instead of “it looks better”

Three numbers make a dark-field setup defensible:

• Defect-to-background contrast at a fixed exposure, measured on a known scratch, not on the best-looking image.
• Background uniformity across the field, because a bright corner will mask defects there. Nine-point sampling is enough.
• Orientation coverage: the same scratch seen at 0°, 45°, 90° and 135° azimuth, with the minimum contrast recorded.

Keep the exposure fixed across candidates. If the comparison is made at “best exposure” for each geometry, the result measures the operator’s patience rather than the illumination.

Combining dark field with measurement

Many cells need both a dimension and a surface check. This is where dark field becomes expensive: a telecentric lens rejects most scattered light, so the image is dim, and pushing exposure to compensate costs throughput. Where the measurement is the primary requirement, the usual arrangement is a separate transmitted channel for the silhouette and a low-angle channel for the surface, each with its own recipe, rather than one compromise illumination. The trade-offs between transmitted and low-angle geometries for metrology are covered in telecentric measurement lighting.

For reflective flat metal where the defect is a marking or a dent rather than a scratch, coaxial illumination is often the better tool, and the selection logic differs: coaxial illumination for reflective metal sets out when that applies.

Hardware that fits the geometry

On the POMEAS side, the geometry is served by low-angle bars and rings. The machine vision lighting range includes bar lights for opposed and four-side grazing, diffuse rings for surfaces that need less directionality, and a four-channel controller that lets each side be balanced independently — which matters, because an unbalanced four-side setup misses one orientation systematically rather than randomly.

When the part is a turned or ground cylinder, a ring at low angle keeps the geometry symmetric as the part rotates, so brightness does not pulse once per revolution. That is a mechanical advantage worth more than a small gain in peak output.

Frequently asked questions

Can I find scratches with a normal ring light if I turn the brightness down?

No. Brightness is not geometry. Dimming a high-angle ring keeps the illumination inside the specular wedge, so flat areas still return light and contrast stays low. The source has to move outside that wedge — physically, not electrically.

How low should the angle be?

Start at 20–30° when the part has tall features, and 10–15° when the surface is flat and the scratches are fine. Then check the shadow length (about h / tanθ) against the spacing of features, and verify with a reference scratch rotated through four orientations.

Why does the whole image brighten on some parts?

Because the surface is diffuse rather than specular: the roughness scatters light in all directions, including into the lens. Dark field cannot separate a defect from a background that scatters the same way; those parts need a different strategy.

Do I need four lights or are two enough?

Two opposed bars cover most orientations but leave a residual blind zone at 90° to their axis. Four sides, or a four-quadrant ring, remove the preference. If the scratch population is known to run in one direction, one bar plus a rotated validation is defensible.

Does a telecentric lens help with dark field?

It helps measurement, not surface contrast. A telecentric stop rejects much of the scattered light, so dark-field images are dimmer. Use telecentric optics when dimensions matter and accept the light cost, or split the two functions into separate channels.

What to send for a lighting review

Send one good part and one part with representative scratches, the surface finish (Ra or process), the field of view, the stand-off available for a low-angle source, and the acceptance criterion in words. A photograph of the current image is useful, but only if the illumination geometry and exposure are stated — otherwise it describes a symptom without its cause.