Flat glass scratch inspection lighting: dark field first

A scratch on clear flat glass is almost invisible under ordinary light: the surface mirrors the room, the scratch mirrors the room, and the camera sees one uniform grey pane. The fix is geometric, not a brighter lamp. Dark-field lighting — bars mounted flush with the pane, light thrown across the surface — turns the flat glass black and every scratch into a bright line that a simple threshold catches. This guide gives the geometry and angles that work, the pixel arithmetic for a 30 µm scratch, the coaxial and backlight channels that complete the station, and an eight-step bench procedure. It is written for anyone building a first glass inspection station who needs defects to show before any software tuning is worth doing.

Which light makes scratches in glass visible?

Dark-field illumination is the primary choice: bar light that meets the pane at less than 20° above the surface reflects off the flat glass and misses the lens, while scratch walls scatter part of the same light upward, so every defect prints as a bright line on a black background.

The law of reflection does the work: a polished flat surface sends grazing light onward at the same low angle, but the steep walls of a scratch turn part of it toward a camera that looks straight down. That law is what separates bright field from dark field lighting, and it makes glass the friendliest dark-field subject there is — an undamaged pane returns almost nothing. Production sheet-glass scanners use this side-incident channel for the same reason, and published setups report catching scratches 3-10 µm wide, far below what a silhouette or an ordinary top light can show.

Two weaknesses come with the technique, and both are handled below: grazing light cannot tell how deep a defect runs, and it lights dust and fingerprints as brightly as real scratches.

Where should the bar light sit for grazing dark field?

Lay the bar lights flat, one on each of two opposite sides of the pane, with their output faces level with the glass surface; if the mechanics force a tilt, keep the beam centre within 20° of horizontal, because a steeper beam fills the background with glare.

For a tilted bar, θ is the grazing angle, h the height of the output face centre above the glass plane and d the horizontal set-back. A face centred 16 mm high at a 60 mm set-back aims at arctan(16 / 60) ≈ 15°. Raise it to 40 mm at the same set-back and the angle reaches arctan(40 / 60) ≈ 34°, steep enough that the flat top surface bounces light straight into a downward-looking lens. Design to the first case.

Two bars cover the strip between them; four bars on all sides cover a pane in a single frame. Bar lights come in lit lengths from under 50 mm to more than 1 m, so match the bar to the pane edge instead of bending the geometry. A low-profile, high-uniformity bar suits a bench station:

The LS-BAR-0484 is a 312 × 16 mm high-uniformity bar light that runs on 24 V. Sit it directly on the bench so its face is level with the glass, aimed across the pane.

When is a coaxial light the better choice?

Move to a coaxial light when grazing light under-lights the defect: the beam travels down the camera axis, the flat pane reflects it straight back as a bright background, and shallow digs and micro-scratches print as dark lines at the same scale as the dark-field image.

On glass this is bright field with inverted contrast: mirror surface bright, broken surface dark. It earns its place on polished or coated display glass, where the finest digs sit too shallow for grazing light to reach. Run both channels on one station — dark field for surface scratches, coaxial for the faint ones — and inspect two images per part. Most coaxial lights cover fields under 200 mm, which matches the small regions where faint digs matter. If a coating ghosts the background, add crossed polarizers, the standard fix for reducing surface glare by cross polarization.

The LS-CXL-0027 is a 60 × 60 mm coaxial light, 4.4 W in red and 7.2 W in white. Size its lit area to the field of view from the pixel math below, not to the whole pane.

Do through-cracks need a different light?

Yes: dark field cannot tell a deep scratch from one that breaks through, so add a transmission channel — a flat backlight under the pane — and image the silhouette, where any opening through the glass shows as let-through light two pixels wide or more.

Surface and silhouette are the two questions in glass inspection, and choosing the lighting angle by surface or silhouette is what splits them. A diffuse backlight answers whether an opening exists; a collimated unit sharpens edges when you must measure a chip's position through the thickness — the trade-off set out for collimated versus diffuse backlights in edge gauging. Keep a 5-10 mm air gap under the pane so a cracked sheet never rests on the emitter.

A 150 × 150 mm lit area covers a pane of the same size in one frame:

The LS-BKL-1468 is a red 150 × 150 mm flat backlight on 24 V. Strobe it with the bars off so each frame sees one channel only.

How many pixels does a 30 µm scratch need?

Budget three pixels across the narrowest scratch you must catch: a 30 µm scratch then sets an image scale of 10 µm per pixel, and a 12 MP sensor with 4,096 pixels covers a 41 mm field at that scale — any coarser and the scratch averages into the background.

Worked example: scale = 30 µm / 3 = 10 µm per pixel. A 4,096 × 3,000 sensor covers 4,096 × 10 µm = 41 mm. Its sensor width is 4,096 × 3.45 µm = 14.1 mm, so m = 14.1 / 41 ≈ 0.34; at a 100 mm working distance, f = 100 mm × 0.34 / (1 + 0.34) ≈ 25 mm, so a 25 mm lens. Start at f/4 — the surface is flat and depth of field is not the constraint — fire the bars 500 µs per strobe and raise exposure until the black background sits at 10-20% of full scale. Dark field returns little light, so the strobe does the work, not the f-number.

A 100 mm pane breaks that budget: 100 mm / 4,096 = 24.4 µm per pixel, and the 3-pixel rule lifts the smallest reliable catch to about 73 µm. Tile the scan at 10 µm per pixel or move to a higher-resolution sensor.

What is the setup procedure, step by step?

Build the station in this order — enclosure, glass handling, light geometry, camera last — because dark-field setups fail most often through room light and a dirty pane long before any optical choice matters.

1. Seal the optics from room light: a black enclosure or a dark bay. The background must sit near the sensor's dark level.
2. Clean the pane with lint-free wipes, and again after every handling step.
3. Lay the glass on 1-2 mm spacers so the bar faces sit level with the top surface, one bar per opposite side.
4. Mount the camera overhead; set focal length, working distance and f-number from the math above, and focus on the top surface.
5. Strobe the bars at 500 µs and raise exposure until the background reads 10-20% of full scale. Strobing also spares the LEDs — why strobe mode makes lights last longer.
6. Threshold at the background plus five sigma of sensor noise; the dynamic-threshold method used for scratch detection holds up better across part-to-part drift than a fixed cut.
7. Confirm on a reference pane: scribe one scratch at the minimum width and one finer; the station must catch the first and miss the second.
8. Add the coaxial or backlight channel and repeat steps 5-7 on its images.

Why does the image look grey instead of black?

A grey background means stray light reached the sensor, and the causes rank in this order: room light leaking past the enclosure, ceiling fixtures flickering into the frame, a glossy fixture surface bouncing specular light up the axis, and fingerprints or dust scattering as bright as real scratches.

Work the list in order. Re-seal the enclosure and recheck. If banding appears, match exposure and strobe to the mains period as explained for camera flicker under artificial lighting. Paint or replace shiny fixture surfaces with matte black. Clean the pane. Flat-sheet surface inspection lives on this discipline — the same controlled lighting rules used for plywood flatness inspection apply unchanged to glass.

When should you not use dark field on glass?

Skip grazing dark field when the defect is not a surface one: bubbles and inclusions inside the glass barely scatter grazing light, frosted or textured glass scatters everywhere so no black background exists, and curved or bevelled zones aim the bars' light straight into the lens.

Decide per defect. Internal inclusions: use the backlight channel — an inclusion blocks transmitted light, so it silhouettes as a dark spot. Frosted glass: switch to coaxial bright field and threshold on local contrast, because the background is bright by nature. Bevels and chamfers: stop the inspected field short of the bevel or mask the bars' inner edge, and give the bevel its own station. When one part mixes defect types, the mapping of lighting selection by inspection task sorts geometry to defect.

For a clear float-glass pane with surface scratches the decision is settled: grazing bar lights on two opposite sides within 20° of horizontal, an optional coaxial channel for faint digs, a backlight for through-breaks, 10 µm per pixel, strobe inside an enclosure. Size the parts with the light finder or start at the lighting step of the configurator; for an unusual pane, send the drawings through the contact page and we will help spec the channels.