Prism vs trilinear line scan cameras for free-fall sorting

Sorting fruit, beans or grain in free fall runs on color: a stem scar, a bruise rim, a pebble among green beans. Whether that color is trustworthy comes down to prism vs trilinear. A trilinear line scan camera lays R, G and B rows along the scan direction on one sensor and recombines them with a fixed line shift. A prism camera splits each captured line inside a dichroic block, so every channel sees the same object point at the same instant. On tumbling product the trilinear shift breaks and colored rims appear — the halo effect — right where your classifier looks. Here are the registration numbers, a worked lane, and the decision for each motion case.

The decision, up front

Take the prism camera when object speed, position or orientation is unknown, which is the definition of free fall. Take the trilinear camera when motion is encoder-locked and steady to within a couple of percent. Take a mono camera with sequenced RGB line lights when you need bands no filter set covers. Those three sentences settle most sorting projects:

• Free-fall or cascading product — produce, beans, grain, recycled flake: prism, RGB or RGB plus NIR. Registration is a property of the optics and does not depend on speed at all.
• Product on a belt or web with an encoder, speed steady within 1–2%: trilinear. The residual error is two lines times the speed error, which is 0.02 lines at 1%.
• Constant speed plus unusual bands — UV, SWIR, two reds for variety separation: mono sensor under sequenced lights, accepting the same staggered capture trilinear has.

If your actual question is prism versus a Bayer area-scan camera, that is a different comparison, covered in prism 3-sensor color cameras vs Bayer. New to line scan entirely? Start with area scan camera line scan mode for moving-web inspection.

How each architecture captures color

A trilinear sensor carries three rows on one die, spaced along the scan direction — the direction the image builds line by line. The gap between neighboring rows is one pixel or less on current sensors. As an object passes, row R samples it, then row G one line period later, then row B. The camera's spatial compensation delays the earlier channels by the row spacing, measured in lines, so the three samples of one object point land on the same image pixel. That shift is exact for one motion only: straight-line travel at exactly the speed the line rate assumes.

A prism camera bonds two to four sensors to one dichroic prism block. Every ray entering the lens follows a single optical path, and the coatings split it by wavelength inside the block, so R, G and B (and NIR on a four-channel block) are taken from the same line at the same instant, each on its own sensor. Tumble, spin and speed changes move all channels together. The row geometry behind both families is laid out in monochrome and colour line scan camera sensor types explained.

The numbers: line rate, row spacing, fringe width

Four relations drive everything that follows:

Row spacing in lines is the distance between the outermost color rows: two line periods for an RGB trilinear with one-pixel gaps, three for an RGB+NIR quadlinear. The fringe width is the object-space distance the product travels between the outermost channels — the width of the band on every edge where the channels disagree about object versus background. Compensation removes it only for pure translation at the speed you set. The leftover from speed spread is fringe_width times the speed error fraction, and cross-scan or rotating motion is not removable at all, because it acts along the row, not along the scan. Setting the line rate itself is its own task: see how to set the line rate of a line scan camera.

Worked example: a 250 mm bean lane at 2 m/s

Green beans fall past a single camera view. Lane width 250 mm, a 2048-pixel sensor across it, fall speed about 2 m/s at the inspection point:

• Object pixel: 250 mm ÷ 2048 = 0.122 mm, so 122 μm per pixel.
• Square-pixel line rate: 2000 mm/s ÷ 0.122 mm = 16,393 lines/s, call it 16.4 kHz.
• Light budget: white light in a short chute collects enough photons at about 6 kHz, so you run 6 kHz and accept a scan-direction pixel of 2000 mm/s ÷ 6000 = 333 μm, a 2.7:1 stretch most sorters tolerate.
• Trilinear R-to-B delay at 6 kHz: 2 ÷ 6000 s = 333 μs.
• Fringe width: 2 m/s × 333 μs = 667 μm, about 5.5 pixels of the 122 μm row.

Every bean edge now carries a 0.67 mm band where R sees bean and B sees chute, or the reverse. On a 12 mm bean that is 5% of the width, per edge, per frame. A bean arriving 15% faster than the dialed-in speed leaves 0.15 × 667 μm ≈ 100 μm of uncompensated shift on every edge. A bean spinning at 2 rev/s moves a surface point 5 mm from its axis at 2π × 2 × 5 ≈ 63 mm/s sideways, and no line shift corrects sideways.

Run the same lane on a prism camera and the three channels share one 167 μs exposure. Motion blur within that exposure is 2 m/s × 167 μs = 333 μm, identical in R, G and B, so edges come out soft but gray. Interior color stays measured and the classifier sees no rim.

The third route costs line rate instead of optics. Sequencing R, G and B light pulses on a mono sensor needs three lines per color set, 3 × 16.4 kHz = 49.2 kHz at square pixels. The LS-LCM-0016 prints a 58 kHz line rate, 7 μm pixels, TDI and a 2K row. How the channels are built that way is covered in color from a mono camera by RGB light sequencing.

Side by side: trilinear, prism, sequenced light

Picking hardware per lane, three catalog points span the range — a 4K color lane, a 16K color lane, an 8K mono lane:

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

Both single-sensor and prism builds appear among the 50 Color models in the camera finder; have the row layout stated on the quote so the fringe math above is anchored to your actual sensor.

Lights and lenses: what both architectures share

Sorting color needs a white light from line scan lights at least as long as the lane, ideally 10–20% longer so the ends do not vignette. The LS-LSC-0527 prints 300 × 17 mm and 87 W red / 108 W white. That 300 mm length covers a 250 mm lane with margin, and white is the mode a color lane runs in. Wider falls take lights near a metre and beyond, and long housings move to 48 V supplies to hold brightness end to end.

Geometry matters as much as spectrum. A bright-field line light catches surface color; a backlight turns the fall into a silhouette for size and foreign material. The choices are laid out in line scan illumination from bright field to backlight.

On the lens side, row length sets the image circle. 16,384 pixels of 5 μm span 81.9 mm, so a 16K row of 5 μm pixels needs an 82 mm circle and a large-format mount. The MV-CL166-91F2C prints 16K, 5 μm pixels, an 82 mm image circle and F-mount. Filter machine vision lenses by image circle and mount before anything else; glass that covers 11 mm will not cover 57 mm because you wished it did.

How to choose, case by case

• Free-fall produce, beans, grain, recycled flake: prism RGB, adding the NIR channel when foreign material such as wood, plastic or caps must separate from product color. The NIR channel arrives registered too. For the views around the fall path and ejector timing, free fall optical sorting: align views and ejector timing is the companion piece.
• Grain on a belt or vibrating conveyor with an encoder: trilinear holds when speed regulation is tight. The pixel budget per grain is its own calculation, worked in grain sorting with a line scan camera.
• Webs — film, print, nonwoven: trilinear, encoder-locked. Lighting for those defect classes is covered in lighting for fabric defect detection.
• Bands beyond RGB — UV mold fluorescence, SWIR moisture, a dual-red variety split: a mono sensor under sequenced lights. The LS-LCM-0038 prints 900-1700 nm on a 1K SWIR row.
• One camera on a mid-width lane: the MV-CL042-91GC-V2 prints 4K, 7 μm pixels, M42 and a 29 mm image circle, in an isolated version. The isolated build suits mounting close to washdown.

When the usual answer is wrong

Prism is not a free upgrade on every color lane. On an encoder-locked web the trilinear residual is two lines times the speed error: 0.02 lines at 1% regulation, and no classifier will notice. Spending the prism block there buys nothing you can see.

Going the other way, software will not rescue a trilinear in free fall. The error is per object and per tumble phase; there is no global shift that works, and per-object correction would need to know each bean's rotation before imaging it. RGB light sequencing does not escape the physics either — its channels are staggered in time exactly like the rows are, so on tumbling product it fringes for the same reason.

One honest boundary: at 40 kHz line rate, a 2 m/s lane has a trilinear rim of 2 m/s × 2 ÷ 40,000 = 100 μm, under one pixel at 122 μm. Small lanes with very strong light can reach that; most color sorting light budgets cannot, which is why the prism rule stands for free fall.

Checks before ordering

• Row spacing between the outermost color rows, in lines and in sensor pixels. One-pixel gaps are typical; some sensors space rows further, and every fringe number above scales with it.
• The line rate the camera sustains at full color output, and the line rate your light supports at the exposure you need. The fringe shrinks linearly with line rate.
• Trigger plan: free-running lines timed to mean fall speed for chutes, encoder-driven lines for belts and webs. The wiring and the software-versus-hardware split are in line scan camera encoder trigger: software or hardware; the scale setting is encoder ticks per pixel for line scan.
• Image circle and mount against your lens stock. An 82 mm circle means large-format glass.
• Light length at least the lane plus 10–20%, and the supply voltage long housings need.
• The channel set: RGB now, NIR later? A four-channel prism or a sequenced-light plan decided at order time costs less than a second camera later.

Start in the camera finder filtered to Color, or open the configurator at the camera step with your lane width and fall speed. The applications desk will check the registration math with you through the contact page.