Detect bruising and fungus on vegetables with SWIR and NIR

Bruising and fungus on vegetables start as water and sugar changes inside the tissue, and a plain colour camera walks past them until browning reaches the surface. On a sort table running 87 ft/min with a robot picking rejects off the line, the questions are which band to image, how fast the camera must sample, and which hardware reaches that band. This guide gives you the band-by-band decision, the line-rate calculation for your belt, camera and strobed-light options from the catalog, and the handshake that sends each reject to the robot. The plan: strobed near-infrared at 850-940 nm first, SWIR at 1170-1450 nm for early or subsurface bruising, ultraviolet only as a last resort.

The decision: image the water, not the colour

A bruise is broken cells. Sap pools in the damaged tissue and starch converts to sugar; fungal growth does the same in reverse, holding water in the mycelium while pigment develops. Water absorbs strongly in the near infrared around 970 nm and again at 1450 nm, and the C-H bonds in sugars absorb near 1200 nm, so bruised and infected tissue darken against sound flesh in those bands. Visible light sees only pigment that has reached the surface, which is why a colour camera shows rot a week late.

The decision is a two-band build. Run a strobed 850-940 nm near-infrared channel first: a monochrome camera plus an infrared bar light is the simplest hardware that reaches water contrast. Add a SWIR channel at 1170-1450 nm when the bruise you must catch is less than a day old or sits under the skin. Keep a visible channel for gross colour grading, and treat ultraviolet as a last resort, never a starting point.

Peer-reviewed hyperspectral work backs these bands. Early apple bruise detection has been reported with SWIR hyperspectral imaging across the 1000-2500 nm range, strawberry bruise contrast has been reported at 825 nm and neighbouring bands, and Fusarium damage in wheat has been separated in both Vis-NIR and SWIR imagery. Treat 825, 970, 1200 and 1450 nm as starting points; the validation run below settles the final pair for your cultivar. The SWIR vs NIR vs visible comparison, the hyperspectral imaging explainer and the SWIR camera applications guide cover the band trade-offs in depth.

What each band shows, and what reaches it

Use a monochrome sensor for every NIR and SWIR channel: a colour sensor wastes most photons behind its Bayer filter. The near-infrared wavelength selection guide shows how the water bands behave, and SWIR wavelength selection by material maps bands to moisture and skin types. If you do test fluorescence, the NIR fluorescence band-selection guide explains pairing an excitation band with the right camera filter.

Sizing the capture for a belt at 87 ft/min (442 mm/s)

Inputs, each with its source:

• Belt speed: 87 ft/min at the inspection point (your sort table). 87 ft x 304.8 mm/ft = 26,518 mm/min; 26,518 / 60 = 442 mm/s.
• Line pitch along travel: 0.22 mm, chosen to place about 4.5 samples per millimetre, so a 1 mm blemish spans 4-5 pixels. The bruises worth rejecting run 5-20 mm across, so this is the resolution floor; the pixels-per-defect method sets the counting rule.

The rule in words: line rate equals belt speed divided by line pitch; the exposure must fit inside the line period; blur equals speed times exposure.

The arithmetic:

• 442 mm/s / 0.22 mm = 2,009 lines/s, so specify a sustained 2 kHz line rate.
• Line period: 1000 / 2009 = 0.50 ms. Cap exposure at 0.45 ms to leave headroom.
• Blur at that exposure: 442 x 0.45 / 1000 = 0.20 mm, under one 0.22 mm line pitch.

Decision: the capture path must sustain 2,000 lines per second at an exposure of 0.45 ms or less. Above a 0.5 ms exposure, adjacent lines overlap and the pitch you paid for disappears into smear. Strobing shifts the budget onto the LED: a 100 μs pulse, a typical strobe duration, blurs only 442 x 0.1 / 1000 = 0.04 mm, so the flash freezes the fruit, not the shutter. The line-rate setup guide works the same calculation from the camera side, and sorting conveyor camera selection covers pixels, speed and blur together.

An area-scan build inverts the logic: a photoelectric sensor triggers one frame and one strobe pulse per fruit, so the frame rate follows the fruit count rather than the belt, and the blur budget stays the same.

Cameras: three builds that reach these bands

Hyperspectral snapshot, 400-1000 nm

One camera covers colour grading, the 825 nm bruise band and the 970 nm water band in a single device. The LS-HSC-0004 area-scan hyperspectral camera covers 400-1000 nm with 1200 bands over USB3, on a 5 MP C-mount sensor; trigger one cube per fruit and classify the spectrum in software. It is the right first camera when the defect mix is unknown or changes with the season.

SWIR area scan

Standard InGaAs SWIR sensors respond from about 900 to 1700 nm, which reaches the 1170 nm sugar band and the strong 1450 nm water band. The LS-IRC-0029 is a 0.3 MP mono SWIR camera with 15 μm x 15 μm pixels, a 1/1.2-inch sensor, global shutter and C-mount lens mount on GigE. At the 0.22 mm belt pitch, 15 μm pixels sit at a magnification of 15 / 220, about 0.07 — comfortable duty for a C-mount lens. Where you need a taller field or finer detail, the MV-CI013-GS-NN raises the SWIR frame to 1.3 MP on an IMX990 sensor, fanless on GigE with a C-mount and a 1/2-inch sensor.

Line scan across the full belt

For a full-width web with no gaps, the classic layout is a line-scan camera looking across the belt, a line light flashing once per line, and an encoder clocking captures at the 2 kHz figure above. Grain sorters run exactly this arrangement; grain sorting with a line scan camera works the pixel budget for the same kind of moving produce.

One lens warning: C-mount lenses are aligned under visible light and then shift focus in the SWIR. Refocus at the working band or select a lens corrected for it — focus shift across wavelengths explains the cause, and SWIR lens basics lists what changes in the glass.

Strobed lighting across the belt

Match the bar to the winning band. The bar lights catalog includes 850 nm and 940 nm infrared versions, 365-405 nm ultraviolet, and white, red, green and blue, so the validated wavelength exists as a catalogue part. Curved fruit throws specular hot spots, so choose a diffused bar and add a crossed polarizer when waxed skins still glare; on a lit factory floor, a bandpass filter over the LED strips the ambient light that would otherwise flood a short exposure. For the visible grading channel, which LED colour gives the best contrast walks through the colour choice.

Size the lit length to the belt width, not to the fruit. Three catalogue bars bracket the common cases: the LS-BAR-1168 is a 423 x 40 mm red diffused bar on 24 V, the bench unit for the visible channel; the LS-BAR-0295 runs a diffused 1272 x 22 mm strip on 24 V for wide belts; and the LS-BAR-0069 spans a 1097 mm lit length on 24 V.

Pulse the bar only inside the exposure window, through a light controller. Short pulses at a few percent duty keep the LED junctions cool and let you drive the bar brighter than its continuous rating, which is exactly what a 0.45 ms or shorter exposure needs.

Triggering and the robot handoff

Track the fruit, not the image queue. A photoelectric sensor at the in-feed fires the frame and the strobe; an encoder keeps counting travel after the camera, so the reject position stays valid when the belt speeds up or slows down — the encoder ticks per pixel guide sets that scale. The result leaves the vision system as a digital reject pulse or an Ethernet message carrying the belt position to the cell controller. At 442 mm/s, a gripper mounted 500 mm downstream has 500 / 442, about 1.1 s, to act; give the cell a default-reject timeout so an uninspected fruit can never slip past in silence. A vision controller with camera ports and industrial I/O closes this loop without a separate PLC, and the robot lifts each marked fruit off between the camera and the packing station.

Prove the band on your own fruit before cutting the frame

1. Collect 40-60 samples: sound fruit, bruised fruit at three ages (fresh, 24 h and 48 h after impact) and visibly mouldy fruit.
2. Image every sample at every candidate band with a fixed exposure, and log the mean grey level in the bruise region and in adjacent sound tissue.
3. Score each band by separation: (mean sound minus mean bruised) divided by the noise σ. Demand 5 or better before trusting a fixed threshold; a band scoring below 3 is not production material.
4. Build with the two best bands: NIR-only if 850-940 nm clears the bar, SWIR added when 1170-1450 nm is the only band above 5.
5. Freeze exposure and thresholds, run a blind pass over unseen fruit, and score false accepts separately from false rejects — the false accept is the one the customer finds.

If fresh-bruise separation at 850-940 nm lands under 3σ while 1450 nm clears 5σ, ship the SWIR channel; do not talk yourself into the smaller build. The spectral classification validation planner turns this sequence into a full protocol with pass criteria.

When the usual answer fails

• Wash water lies. Free water on the skin absorbs at 970 nm and 1450 nm exactly like bruised tissue, so a wet line reads every droplet as damage. Dry with an air knife before the camera, or move the decision to the 1200 nm band, where standing water is far weaker.
• Thermal cameras answer the wrong question. A long-wave infrared camera images emitted 8-14 μm radiation, which is temperature; the respiration warmth of a fresh bruise on cold produce sits far below what such a sensor resolves. LWIR belongs on hot-spot and temperature work, not bruise.
• Internal rot has no optical signature from outside. Mould sealed inside a closed fruit shows nothing in any band; that is a process problem, not a camera problem.
• The season shifts the sugar bands. Cultivar and maturity move the C-H response, so rerun the validation pass each season under the same frozen protocol.

Checks before you order

• Measure the belt speed at the camera, not on the nameplate: the entire 2 kHz and 0.45 ms budget hangs on that one number.
• Measure the belt width and add margin: a bar lit short of the full field pushes the edges below threshold exactly where fruit drifts.
• Choose snapshot-per-fruit or continuous line scan first, because it sets the trigger hardware and decides whether the 0.45 ms cap applies per line or per strobe.
• Check the ambient light over the inspection window: overhead fluorescents flicker at twice the mains frequency and inject noise into short exposures — strobe plus bandpass wins.
• Time the pick window: camera-to-gripper distance at 442 mm/s fixes how fast the handshake must be.

Parts list and next step

In shopping order: the camera that matches the validated band (hyperspectral cameras for the 400-1000 nm snapshot build, infrared cameras for the SWIR channel), a C-mount lens focused at the working band, the strobed bar lights for the winning wavelengths, a light controller, the photoelectric sensor and encoder, and a vision controller to close the loop to the robot. The camera finder and light finder filter by resolution, interface and wavelength, and the light configurator assembles bar and controller in one pass. For a quotation across the whole kit, the quote page handles multi-line requests — or contact us with your belt speed, belt width and fruit list, and we will size the build with you.