Camera settings for steel coil defect detection: dial in sharpness

Steel coil is the hardest surface in the plant to image sharply: mirror-bright, fast-moving, and judged by defect lists written in fractions of a millimetre. The settings that deliver sharp frames are decided long before the menus, by three measurements: the smallest defect in the quality spec, the strip width and the line speed. Size pixels to the defect first, exposure to motion second, aperture to depth of focus third, and only then touch gain and lighting. This guide works a 1,250 mm wide cold-rolled line at 0.5 m/s end to end, with the arithmetic shown, so you can rescale every number to your own strip in minutes.

How many pixels do you need across a steel coil?

Give the smallest defect your spec names at least 3 pixels across: for a 0.3 mm scratch that is 0.1 mm per pixel, and 12,500 pixels to span a 1,250 mm strip.

Three pixels detect a feature reliably; five or more measure it. Below three, the defect hides in surface grain and sensor noise. Above the rule, you buy bandwidth and light for nothing. The arithmetic runs backwards from the spec:

Twelve thousand pixels across one strip is why production coil inspection uses several cameras per surface rather than one. A 12 MP sensor in the 1.1-inch class carries 4,096 × 3,000 active pixels, so one lane images 4,096 × 0.1 mm = 409 mm of width, and four lanes cover 1,636 mm — a 1,250 mm strip with overlap at every seam. The pixel rule for defect detection gives the sampling argument in full. Size your own variant here:

Does shutter type matter when the coil is moving?

A moving coil needs a global shutter: a rolling shutter scans rows over tens of milliseconds, so at 0.5 m/s the bottom of the frame images the strip 15-20 mm further along than the top and every vertical edge leans.

Rolling shutter is a readout property, not a menu setting — no exposure value repairs it. A 12 MP rolling sensor reads a full frame in roughly 30-40 ms; at 500 mm/s that is 15-20 mm of smear between first and last row. Welds shear, edges lean, width measurements drift. A global shutter exposes every row at the same instant, so a 200 µs frame freezes the whole strip at once.

Rolling shutter stays the right tool where the strip stops. Master images for the defect library, training sets and offline tuning reward the extra pixels of a mono sensor. For that stopped-strip work, pick resolution over shutter speed:

The moving-line cameras come from the area scan cameras range; filter global shutter first, resolution second.

What exposure time freezes coil motion?

Set exposure from motion blur, t_max = pixel pitch ÷ line speed: at 0.1 mm per pixel and 0.5 m/s the ceiling is 200 µs, and a half-pixel blur budget tightens it to 100 µs.

One pixel of blur keeps edges inside the 3-pixel budget. Frame rate follows from coverage, not the other way round: each frame images 3,000 px × 0.1 mm = 300 mm along the strip, so with 10% overlap the camera fires every 270 mm of travel — 500 ÷ 270 ≈ 2 fps. Every GigE or USB3 camera of this class runs ten times faster, which is your headroom for triggers and re-shots.

Two habits protect the 200 µs figure. Turn auto exposure off, or cap it, because the camera will trade shutter speed for brightness at the first dark weld (where the cap setting lives). And let the strobe, not the factory ceiling, make the image: 200 µs is 2% of a 50 Hz mains half-cycle, so ambient flicker cannot band the frames (the banding mechanism, diagnosed).

Which f-number keeps a steel strip sharp?

Work between F4 and F5.6: at F4 the diffraction spot is 5.4 µm, under two 3.45 µm pixels, and F5.6 still holds about ±34 mm of depth of field — enough for a ±15 mm passline swing with margin.

At F8 the spot reaches 10.7 µm — three pixels of blur before anything else moves — so F8 is the ceiling, not a setting to grow into. Depth of field is generous at this magnification: 2 × 5.6 × 6.9 µm × 1.0344 ÷ 0.0344² ≈ 67 mm total, about ±34 mm either side of the working distance. That swallows coil crown, weld bumps and roll rebuilds.

Focal length follows from magnification, m = sensor width ÷ field of view = 14.1 mm ÷ 409 mm = 0.0344, through f = WD × m ÷ (1 + m). A 600 mm working distance asks for a 20 mm lens; the stock steps on either side are 16 mm at WD 480 mm and 25 mm at WD 750 mm, and both hold 0.1 mm per pixel. Take the 16 mm route and the working distance drops under half a metre, which also trims shimmer over warm strip. The LS-FFL-0567 covers it: 16 mm focal length for a 1.1-inch sensor at 20 MP, F2.8-22, under 0.3% distortion.

The lens finder filters focal length and image circle; the lenses range lists every mount and format.

How much gain can you use before noise wins?

Keep gain at 0 dB and buy light instead; if the histogram still sits below half scale at full strobe current, take up to +6 dB and stop, because beyond that noise eats the faint stains you are hunting.

The attack order is fixed: exposure time up to the 200 µs budget, then aperture toward F4, then strobe current, then gain. Analog gain multiplies sensor noise along with signal, and low-contrast defects — shallow stains, light oxidation — live a few grey levels above the background. Six decibels doubles the signal for a modest noise penalty; twelve starts to bury them. Mono sensors earn their keep here: each Bayer-filtered color pixel collects roughly a third of the photons its mono neighbor sees, so mono returns about 1.7 times the signal-to-noise ratio at the same exposure (shot noise scales with the square root of collected photons). Set exposure, aperture and gain manually; the exposure and gain balance guide walks the trade-off.

Which lighting angle shows which defect?

Shoot raised defects — scratches, dents, slivers, edge burrs — with low-angle dark-field bars at 10-20° grazing, and discoloration, rust and oil stains with diffused bright field; one station usually needs both banks.

Dark field and bright field are not rivals; they are two banks pointed at two defect families. Raised metal scatters a grazing beam, so bars along both strip edges light defects bright against a dark mirror. Color changes need even, diffuse light from above, or the specular sheen of the strip swamps them. The lighting geometry guide lays out the angles; low-angle lighting on metal and scratch detection with dynamic thresholding show each family on real parts.

Strobe both banks. A 200 µs pulse at 2 fps is 0.08% duty — 0.4 ms of light per 500 ms — so the bars run cold and can be driven above their continuous current for the pulse; the controller's duty-cycle rating sets the ceiling. White LEDs serve both defect families, and a crossed polarizer over the bright-field bank calms any surviving sheen (cross polarization). Pick lengths in the bar lights range via the light finder, and let the lighting configurator pair them with a strobe-capable controller.

One bar per lane per family, strobed:

What does a full-width 1,250 mm station add up to?

Four 12 MP global-shutter cameras at 0.1 mm per pixel cover 1,636 mm of strip, so a 1,250 mm coil needs four lanes; each runs about 2 fps at 200 µs and F5.6.

The whole design in one table:

The catalog covers all three duties in that table:

All 3 share: Gain 0 dB to 24 dB; Binning Supports 1 × 1, 2 × 2, 4 × 4; Decimation Supports 1 × 1, 2 × 2, 4 × 4; Lens mount C-mount; Compatibility USB3 Vision, GenICam; Certification CE, RoHS, KC. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The lanes take the global-shutter models, the 20 MP mono shoots stopped-strip masters, and the 6 MP mono serves a narrow static lane or a prototype bench. Choose lane hardware with the camera finder — global shutter first, then resolution — and let the camera configurator match mounts and cables.

What should you check before ordering?

Before you order, confirm three numbers on paper: pixels across ≥ 12,500, exposure ≤ 200 µs, and depth of field ≥ ±15 mm; then add one spare lane and a controller that fires every light bank from one encoder pulse.

The rest of the list is physical, not optical:

• Defect spec in millimetres, signed by quality — every pixel count hangs on it.
• Line speed range including jog and rollback, not just nominal speed.
• Passline swing measured over a full shift of coil diameters, not one coil.
• Sensor diagonal inside the lens image circle; no stray spacer rings on the C mount (the C-mount spacing trap).
• Cable runs: USB3 for the few metres to a local node, GigE for the 100 m run back to the cabinet.
• Strip temperature: above roughly 80 °C, convection shimmer across a 480-750 mm working distance starts to bend the imaging path — shorten it or shield it.
• Focus proof per lane: image a scratched scrap plate, maximize a focus metric, then lock the barrels.

For how banks mount along a real coil line, the steel coil end face layout shows the multi-camera arrangement on a comparable line; the lens configurator then pairs image circle to sensor.

When should you switch to line scan instead?

Switch to line scan when speed or width breaks the area-scan math: above roughly 1 m/s, or past about 30,000 pixels of demand, two 16k line-scan sensors beat a bank of area-scan lanes on cost, sync and optics.

Line scan replaces lane planning with one sensor line and an encoder: no overlap seams, no lane-to-lane shading match, uniform illumination along the full width. Production coil classifiers are built exactly that way, sampling 0.1-0.5 mm per pixel on both surfaces at once. Area scan keeps the win for pilot lines, narrow strip, end-face and edge stations, and anywhere the line stops. The trigger moment is visible in the budget itself: when pixel demand passes about 30,000 across, or speed passes 1 m/s and the frame-rate table stops closing, move the arithmetic to the line rate calculator. The same blur-and-pixel logic for transport belts is in conveyor camera selection, and the applications topic page collects the full-width layouts.

Ready to size it? Send the strip spec, defect list and speed profile through the quote request for a lane-by-lane sizing, or the contact page for a quick check first.