Detecting dents and pits in metal surfaces with low-angle lighting

A dent or a pit is a geometry defect: the metal is all there, just lower than the surface around it. Treat one like a print defect and the inspection never settles — pits slip through on one shift, grain speckle becomes a false call on the next. The fix is lighting geometry, not more software. This guide is for anyone setting up camera inspection on flat or gently curved metal: stamped panels, machined faces, extrusions, rolled strip. You will set the light angle, budget pixels against the smallest pit, build an opposed pair of bar lights, and verify the image with numbers. The recipe up front: light the surface at 15-30 degrees from horizontal, with at least 3-4 pixels across the smallest pit.

Why flat light loses dents and pits

Under even, on-axis light, a flat metal face returns nearly the same gray everywhere. A crater 0.05 mm deep changes that return by a handful of gray levels — inside the noise of texture, mill finish and illumination drift. That drift is the instability the inspection has been fighting.

Tilt the same light to a grazing angle and the crater becomes the brightest thing in the frame. The wall facing the light reflects the beam up into the lens; the opposite wall blocks it and throws a shadow. One pit now reads as a bright crescent against a dark wedge, a signature no grain streak produces.

A pit 0.5 mm across and 0.05 mm deep has walls tilted about arctan(0.2) ≈ 11 degrees. At grazing incidence that slope redirects the light completely; under on-axis light it shifts brightness by a few percent. That bright-versus-dark behavior is dark-field lighting in the illumination geometry families.

Lighting for metal surface defect inspection runs the same reasoning across scratches, stains and dents; this guide stays on the dent case and works the numbers.

The two numbers that decide the set-up

Elevation: 15-30 degrees measured from the part surface. Start at 20. Below about 15 degrees the bars swing far out from the field — the offset grows as the angle flattens — and every burr throws a long shadow across real features. Above about 30 degrees the sensitivity fades: the gap in illumination angle between the flat field and an 11 degree pit wall narrows, so shallow pits sink back toward the background.

Pixels: at least 3 across the smallest pit, 4-5 to grade depth. A blob detector needs roughly two pixels to see anything and several more to tell shape from noise. Shallow pits are low-contrast targets, so they need the margin. Work the floor backwards:

A 2448 px camera over a 330 mm field resolves 0.135 mm per pixel, so the floor sits near 3 × 0.135 ≈ 0.40 mm. The calculator below runs it for your camera and field:

Worked example: 0.5 mm pits on a 300 mm brushed aluminum panel

The inputs are a typical stamped-panel case; substitute your part's numbers and follow the same steps.

• Part: 300 mm × 200 mm brushed aluminum stamping, inspected in one shot.
• Smallest defect in the spec: pits Ø0.5 mm and up, about 0.05 mm deep.
• Camera: 2448 × 2048 px (5 MP class), 8-bit gray, viewing straight down.
• Lighting: opposed bar pair, 24 V, target elevation 20 degrees, bars 50 mm above the surface.
• Working rules: field of view = part size + 10%; pixel floor 3 px, grading target 4 px; lit length at least 1.2 × field of view.

Field of view: 300 × 1.10 = 330 mm on the long side; 200 × 1.10 = 220 mm on the short side.

Pixel scale on the long side: 330 mm / 2448 px = 0.135 mm/px. The 0.5 mm pit spans 0.5 / 0.135 = 3.7 px.

Pixel scale on the short side: 220 mm / 2048 px = 0.107 mm/px. The pit spans 0.5 / 0.107 = 4.7 px.

The worst axis governs: this camera sees every 0.5 mm pit at 3.7 px — past the 3 px detection floor with about 24% margin, short of the 4 px grading target.

Decision: keep the 5 MP camera for go/no-go detection. If pits must be graded, or the spec tightens below 0.40 mm, split the field into two 165 mm shots — 165 / 2448 = 0.067 mm/px, so 7.4 px per pit — or step up to a 12 MP sensor, where 330 / 4096 = 0.081 mm/px gives 6.2 px.

Bar offset: height / tan(elevation) = 50 / tan 20° = 50 / 0.364 ≈ 137 mm from field center on each side. Bar lit length: 1.2 × 330 = 396 mm.

That lit length lands the bar in the 400 mm + size bracket.

The diffused face is deliberate at this angle: a bare LED row would stripe its own reflection across brushed grain, and the diffuser spreads it into an even line. Mount with a gauge, not by eye — the offset and elevation above are the numbers that make or break the signature.

Build the lighting station

You need the camera and lens fixed for the chosen field, two bar lights, a controller with an independent channel per bar, brackets that hold a measured angle, a shroud, known-good parts and parts with a seeded defect — a pressed dent or an EDM notch.

1. Fix the camera square over the part and set the field to part size + 10%. Compute pixels per millimeter and confirm at least 3 px across the smallest pit before touching the lighting.
2. Mount the bars at 20 degrees elevation, offset ≈ height / tan 20° per side, with the light travelling along the grain on brushed parts.
3. Wire each bar to its own controller channel so the two can be strobed independently.
4. Size the strobe pulse to freeze motion: blur = line speed × pulse time, kept under a third of the smallest pit. Sizing the strobe pulse for high speed barcode verification works the same arithmetic line by line.
5. Balance with one bar at a time: the flat field should read below 10% of full scale, with any seeded-pit crescent standing well clear of it.
6. Detect both halves of the signature: threshold the bright crescent and the dark wedge separately, or run scratch defect detection with lighting and dynamic thresholding so the threshold follows the local floor.
7. Verify: 30 known-good parts with zero calls, then every seeded part caught with each bar on its own, repeated across a full shift.
8. Kill ambient light: overhead shop lighting raises the dark floor and eats contrast, so shroud the cell or fit a bandpass filter matched to the bar wavelength.

Between 100 and 400 mm of field the bar lights range is wide. The LS-BAR-1047 is a 290 × 30 mm standard LED bar light on 24 V.

Polished, hard parts scatter little light back — that is when the LS-BAR-0668, a 312 × 36 × 26 mm high-luminance bar light on 24 V, earns its place.

Grazing geometry throws away most of the beam, so the higher drive in the same footprint buys back exposure margin.

Bars, coaxial or dome

Coaxial is bright field by construction: the light travels down the camera axis, so only tilted walls change the return, and shallow craters change it little. Its home ground is the shared station, one shot that reads an engraved or laser mark and catches the deeper dents. For that combined job, the LS-CXL-0027 is a 60 × 60 mm coaxial light rated 4.4 W in red and 7.2 W in white on 24 V.

The coaxial lights line runs from under 50 mm up past 400 mm, and machine vision lighting selection by inspection task sorts these geometries by what the camera must decide.

Shiny curved parts — caps, can ends, bottle shoulders — break a bar's grazing angle, because the angle is only right along one line of the curve. A dome light wraps the whole part in diffuse light from every direction, so the crater still reads, at lower contrast than dark field. Size the dome so the emitting area clears the part width with room to spare. The LS-DOM-0085 is a Ø166 mm dome light with a Ø116 mm emitting area on 24 V.

Keep the part inside about three-quarters of the emitting area: for this dome, roughly 87 mm across. Dome lights for shiny, curved and wrinkled parts covers dome sizing and failure modes in detail.

When the usual answer is wrong

Brushed and textured surfaces. Grain is a field of elongated streaks, and light crossing the grain turns every streak into a tiny crescent — the main source of false calls. Run the bars so light travels along the grain. Pits are craters, so they catch light from any heading and stay visible.

Oiled, wet or mirror-finish parts. Oil films smear the crescent edges, and a near-perfect mirror returns almost nothing to scatter. Wiping stages and wavelength swaps will not fix that physics; move to depth sensing, where surface defect detection with a 3D laser line profiler measures the dent itself. On shiny metal, improving the 3D point cloud is the hard part.

Glinting flat fields. Cross polarization — polarizer on the bar, analyzer on the lens — kills flat-field sheen; copper surface inspection lighting works through that fight. The trade is that polarization dims the crescents too, so budget drive or exposure for it.

Depth questions. If the question is how deep rather than whether, stop thresholding 2D shading: shading moves with every tilt of the part, a height map does not.

Checks before you commit

• Flat field below 10% of full scale with one bar on — the floor must stay dark for the threshold to have headroom.
• Seeded-pit crescent at least 25-30% of the gray range above local background — contrast that survives drift, and why adequate contrast alone is not enough.
• At least 3 px across the smallest pit on the worst axis — computed from your own field and pixel count.
• Every seeded pit caught with each bar on its own — proof the opposed pair covers both slope directions.
• Dark floor moving less than 5 gray levels with shop lights on and off — proof the shroud or filter holds.
• 30 known-good parts across three shifts with zero false calls — stability is a rate over time, not one good hour.
• Elevation and offset taken from a gauge — eyeballed angles slide toward 45 degrees, which is bright field again.

List the part size, the smallest pit and the line speed, and the rest is arithmetic: the light finder filters the bar, coaxial and dome ranges by size and wavelength, the lighting configurator pairs lights with controller channels, and the lighting topic guides cover the neighboring geometries. For a part that resists all of the above, bring a photo and the spec to contact the team.