Lighting for metal surface defect inspection

A machined metal part rarely fails inspection for one reason only. The same stamped bracket can carry rust spots from storage, scratches from handling and a flatness error from the forming press, and each defect answers to a different lighting geometry. This guide gives the two-channel set-up to start with on flat steel and aluminium parts: a blue coaxial light for rust and flatness, a low-angle dark-field bar light for scratches, one strobe pulse each and two images per part. It works through the contrast argument defect by defect, a worked example that sizes the camera from the smallest rust spot you must catch, and the bench checks to run before ordering hardware — written for the production or quality engineer specifying the station.

The decision: one geometry per defect

Rust, scratches and warp each disturb light in a different way. A rust spot is an oxide patch that absorbs at the blue end of the spectrum while the steel around it reflects; a scratch is a groove whose walls scatter light sideways; warp tilts the whole reflected field. No single light maximises all three contrasts at once, so the working answer is two channels fired in sequence.

The rows follow the surface-versus-silhouette logic in machine vision lighting angles: surface or silhouette?, and each geometry is one family from the guide to machine vision illumination types. If you buy one light first, buy the coaxial: it covers two of the three rows and stays useful as an alignment light while the dark-field channel is being set.

Rust spots: blue coaxial light

A coaxial light feeds its LEDs through a beamsplitter, so the light travels down the lens axis and returns along it off a flat surface. On ground or rolled steel the whole face comes back bright and even. An oxide patch absorbs blue instead of returning it, so every rust spot reads dark against a bright background — a plain threshold separates them.

Blue is the colour to pick. Bare steel reflects across the whole visible band, rust does not, and the blue end is where the gap is widest; red does the opposite and washes rust into the background. Blue also sits far from the amber cast of typical shop lighting, so less ambient leaks into a short exposure. The LS-CXL-0227 is a 50 x 50 mm blue coaxial light wired for 24 V, and its emitting face covers a bracket-sized field with margin for placement tolerance.

To choose colour from data rather than argument, photograph the same sample under two colours and compare — the method is in which LED colour gives the best contrast. On polished or plated faces, glare can saturate the image before rust shows; cross-polarising this channel fixes it, as set out in the cross polarization guide.

Scratches: low-angle dark field

A flat surface acts as a mirror. Light that arrives at a grazing angle leaves at a grazing angle and misses the lens, so the background stays black. A scratch breaks the mirror: its walls catch the light and scatter part of it back up the lens, and every scratch glows as a bright line on black. The geometry does the work, not raw brightness.

Mount the bar nearly parallel to the part plane, tilted 10-20° up from the surface and 50-100 mm away, aimed across the inspected face. At grazing incidence only a small fraction of the emitted light scatters back to the sensor, so shallow scratches need intensity at the surface — this is where high-luminance bars earn their name. The LS-BAR-0400 is a 312 x 30 mm high-luminance bar light on 24 V.

For a bracket-sized face the realistic choice sits between three bars in this class: the LS-BAR-0400 and the LS-BAR-0668 high-luminance bars at 312 x 30 mm and 312 x 36 x 26 mm, and the LS-BAR-1047 standard bar at 290 x 30 mm, all 24 V:

Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Keep the standard bar for deep, coarse scoring, where scatter is strong and surface brightness is not the limit. What the software does with the bright-line image afterwards — thresholds, dynamic models, classification — is worked through in scratch defect detection with lighting and dynamic thresholding.

Flatness and warp: uniform bright field

A flat face under an even field reflects evenly; a warped face does not. Bends, dents and camber show as shading gradients across the image, and a quick bench confirmation is to tilt the part slightly and watch the gradient move. The requirement is uniformity over the whole face, which makes a diffused bar the simplest source: its output is spread into a broad, even field rather than a directed beam. The LS-BAR-1276 is a 372 x 28 mm white diffused bar light on 24 V, long enough to cover a bracket-sized field with room for two-up fixturing.

Two limits belong in the open. Shading finds warp qualitatively; a drawing that states a flatness tolerance in tenths of a millimetre needs a measurement, not a picture, and surface defect detection with a 3D laser line profiler shows that path. And keeping a field even across 400 mm or more is its own balancing problem, covered in large field of view lighting. On smaller faces the coaxial channel already bought for rust does this job, so a two-channel station needs no third light for flatness.

Worked example: resolving 0.15 mm rust spots on a 120 x 20 mm slot-cover bracket

The inputs:

• Part: a flat slot-cover bracket, taken as 120 x 20 mm — the part class this kind of station is built for. Add 5 mm of margin on every side and the field of view is 130 x 30 mm.
• Acceptance limit: the smallest rust spot to catch is 0.15 mm across; set your own value from your acceptance criteria before you size anything.
• Camera: a 5 MP sensor, 2592 x 1944 pixels, a common class for this work.
• Rule: a feature needs about 3 pixels across to detect reliably — the usual working minimum in machine vision.

The rule in words: the pixel scale is the field width divided by the sensor width in pixels, and the smallest detectable feature is three times that scale.

The arithmetic:

• Pixel scale: 130 mm / 2592 px = 0.0502 mm/px — call it 0.05 mm/px.
• Smallest spot: 3 px × 0.0502 mm/px = 0.15 mm.
• The height direction has more margin: 30 mm / 1944 px = 0.0154 mm/px, so the field width governs.

Decision: one shot per part with a 5 MP camera catches rust spots down to about 0.15 mm on this bracket. That meets a 0.15 mm limit with no slack, and it fails a 0.10 mm limit. For 0.10 mm the scale must reach 0.10 mm / 3 px = 0.033 mm/px, which takes 130 mm / 0.033 mm/px = about 3,900 px across the field; a 4096-pixel sensor gives 130 mm / 4096 px = 0.032 mm/px, so 3 px = 0.095 mm. Keeping the 5 MP camera and splitting the part into two 65 mm passes works too: 65 mm / 2592 px = 0.025 mm/px, so 3 px = 0.075 mm, at the cost of two images per part.

Keep one distinction in mind: lighting and sampling are separate limits. The blue coaxial channel supplies the contrast that makes a spot visible; the arithmetic above decides whether the sensor puts enough pixels on it. If spots below the floor must be found, no light setting will fix it — narrow the field or move to a wider sensor.

Fire the two channels in sequence

Run the channels one after the other, never together. The coaxial channel floods the face along the lens axis, which is exactly the light the dark-field channel needs absent: with the coax firing, the black background lifts and scratch contrast disappears. Each light takes its own short strobe pulse inside the camera exposure — 0.1 to 1 ms is typical at bench working distances — and the camera saves one image per channel while the part sits still. A multi-channel LED light controller switches each channel on its own trigger line from the camera's exposure output; short pulses also keep flicker from shop lighting out of both frames, and camera flicker under artificial lighting walks the diagnosis when banding still appears. Sequencing patterns for three or more channels are covered in sequential machine vision lighting for multi-view inspection.

Set both channels up in the light configurator and filter the lights range by colour, size and voltage.

Checks before ordering

• Image real samples, not one good part. Pull parts at both ends of the finish range — bright-rolled and scaly, scratched and clean. Rust depth and surface finish move contrast more than any calculation here predicts.
• Photograph your acceptance spot. Make a reference of the limit size, an etched coupon or a calibrated defect standard, and confirm it clears threshold under both channels.
• Confirm coverage, not just size. The coaxial emitting face must cover the part diagonal, not just the long dimension.
• Test the dark background with the shop lights on. Fire the bar at working drive and check the background still reads black; if it lifts, shield the station or shorten the pulse.
• Wire one strobe line per channel and prove the sequence. The two channels must never overlap, on the bench or in the machine handshake.
• Bolt the lights to the station frame, not a guard. The grazing angle that shows scratches moves if the bar shifts a few millimetres or a few degrees, so the mount needs a positive stop.

When the usual answer is wrong

Dark finishes break the first row of the table. An oxide patch on black-oxide or anodised steel gives no blue bright-field contrast, because the part and the defect are both dark; move the rust check to grazing light or a controlled dark background instead, on the same reasoning as black O-ring lighting.

Curved and highly polished faces scatter a coaxial field everywhere. Surround the part with diffuse light instead — dome lights for shiny, curved and wrinkled parts sets out that geometry, and lighting shiny metal pins without glare fights the same battle on small hardware.

Rust pits below the worked example's floor are a sampling problem, not a lighting one: narrow the field or change the sensor, and check what the acceptance criteria actually demand before chasing smaller spots — steel plate specifications, for one, publish explicit surface defect acceptance limits for ASME SA-516 plate. And once lighting delivers contrast, the decision logic still has to be proven against golden parts and a scored defect library; the checklist in commissioning an AI vision inspection for defect detection covers that stage.

The set-up to start with

Two channels, two images, one part: a blue coaxial light sized to the part face for rust and flatness, a high-luminance bar at 10-20° above the surface for scratches, both strobed in sequence by one controller. Browse the coaxial lights, bar lights and spot lights ranges — spot lights cover the same jobs on very small fields — or step through the light configurator. More geometries, from backlight to dome, are collected on the lighting topic page. Before committing hardware, run a sample part through both channels and score the grey gap between defect and background; why contrast margin survives production explains the margin that separates a lab pass from a stable station.