SWIR fruit sorting: subsurface defects, bruises and rot

An RGB grader sees colour and shape, so a bruised apple under unbroken skin, a lemon rotting from the inside and a stone riding a raisin stream all pass as good product. Between 1000 nm and 1700 nm, water, sugars and oils absorb light, and a short-wave infrared camera turns those differences into image contrast. This walkthrough is for sorter builders and plant engineers adding SWIR to a fruit, dried fruit or nut line. You will leave with the band, the camera, the lens and the illumination chosen, one lane sized with real arithmetic, and a decision for every crop on the line.

The decision first

Sort thin-skinned fresh fruit — apples, pears, kiwi — on the water band with an InGaAs camera and reject what reads dark. The 1450 nm water absorption peak is the strongest inside an InGaAs sensor's reach, and a bruise is localised water. Add a 1650 nm reference frame and divide: the ratio removes illumination falloff and fruit curvature, leaving water content as the only contrast. For thick-peeled citrus, move to 970–1200 nm, where water absorbs less and light returns from deeper tissue. For dried fruit and nuts you discriminate materials, not water: use two bands, one where the product absorbs and one where it reflects, and reject whatever does not behave like product.

On hardware, commit to this rig for a first line: a 0.3–1.3 MP InGaAs camera with global shutter and C-mount, a lens rated for 900–1700 nm, and a fibre-coupled halogen or water-band LED source behind a diffuser. The vegetable side of the same physics is covered in detecting bruising and fungus on vegetables with SWIR and NIR; here the focus is fruit lines and the numbers that size them.

Why damage reads dark on the water band

Liquid water absorbs in distinct bands near 970 nm, 1200 nm, 1450 nm and 1950 nm. The 1450 nm peak is the strongest of these inside the 900–1700 nm reach of InGaAs; 1950 nm needs a different, cooled detector family. Those peaks are why the band, not a camera trick, does the sorting. The same physics drives near infrared wavelength selection for water and silicon inspection and the moisture curves in SWIR wavelength selection by material.

A bruise is ruptured cells. Sap floods the air spaces between them, so the damaged volume holds more water than sound flesh and absorbs more of the water-band light; sound tissue, full of scattering interfaces, returns bright. The bruise reads dark within hours of impact, through unbroken skin. Soft spots, early rot and mold change moisture the same way, which is why one band covers a whole defect family. The mirror image of that strength: the light reaches only the outer few millimetres of tissue — enough for impact damage and skin-adjacent decay, not deep core browning.

Nuts add a second handle. Kernel oil has C-H overtone absorption near 1200 nm that shell material lacks, which is how an in-shell almond hiding in a kernel stream gets caught. Inorganic contaminants — stones, glass, metal — are spectrally flat where fruit absorbs, so they stand out in the ratio image. Choosing the band by what the target material does is the general method in SWIR camera applications: pick the band by what the material absorbs.

Camera: InGaAs, not silicon

Silicon sensor response fades between 1000 nm and 1100 nm, so no filter turns a standard camera into a water-band camera — the photon has to reach a detector that can convert it, as camera sensor technologies by spectral range lays out. You need InGaAs. Two candidates from infrared cameras:

The LS-IRC-0029 is the classic sorting format: 0.3 MP, 15 µm × 15 µm pixels, 1/1.2" sensor, global shutter, C-mount, GigE. Big pixels collect light fast for short exposures, and global shutter stops smear on rollers.

The MV-CI013-GS-NN brings 1.3 MP on a 1/2" sensor with C-mount and GigE, for lanes that grade surface finish and defects in one image.

If the enclosure runs warm, InGaAs dark current climbs with sensor temperature — TEC sensor cooling in SWIR cameras covers when cooling pays for itself. For where SWIR sits against NIR and visible options, see SWIR vs NIR vs visible imaging: which band to inspect with.

Lens and illumination on the water band

Glass that images crisply in the visible is not corrected in the infrared: focus shifts between bands and contrast drops. Focus shift across wavelengths shows why, and SWIR lens and UV lens basics covers what to buy instead. Three C-mount SWIR lenses from thermal, SWIR and UV lenses that cover the water bands:

All 3 share: Mount C-mount; Focus Manual; Operating temperature -10 to 50 °C; Storage temperature -20 to 60 °C. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Match the image circle to the sensor diagonal first. The sorting sensor's 1/1.2" format needs 1" coverage: that is the LS-SWL-0229, a 35 mm lens at F1.4 transmitting 900–1700 nm. A 1/2" sensor pairs with the LS-SWL-0246 at 25 mm and F1.8, whose 400–1700 nm range also passes visible light for a colour camera sharing the view through a splitter. The 2/3" LS-SWL-0252 at 35 mm and F1.4 sits between them.

Two sources work in this band. LED arrays emit on the water absorption peak directly, run cool and stay put. A fibre-coupled halogen source with water-peak and 1650 nm bandpass filters (25–50 nm wide is the usual choice) lets you re-tune the band per crop by swapping filters — build this rig first, because it is also the bench-test rig below. Fruit skins are waxy curves: put a diffuser in front of either source, or specular highlights saturate and drown the bruise contrast. The filter-first method is laid out in near IR bandpass filter contrast: test IR before you buy.

Sizing one lane: apples at 0.5 m/s

Take the classic sorting sensor: 640 × 512 InGaAs pixels at 15 µm pitch, a 9.60 mm × 7.68 mm active area. The lane carries 80 mm apples on cups, 320 mm across the belt, at 500 mm/s.

Target 0.5 mm per pixel so a 2 mm bruise spans 4 pixels — above the 3-pixel minimum for stable detection. The sensor width then fixes the geometry: m = 9.6 mm / 320 mm = 0.03. With a 35 mm lens, WD = 35 mm × 1.03 / 0.03 = 1.20 m, and the vertical field is 7.68 mm / 0.03 = 256 mm, so one apple fills a third of the frame with room for the cup. The LS-IRC-0029 is this class of camera.

Motion sets the exposure: at 500 mm/s, 1 ms smears 0.5 mm — one pixel. Hold exposure at or under 1 ms, or strobe the halogen and freeze the fruit. InGaAs sensitivity with an F1.4 lens and a diffused panel reaches that comfortably; if it does not, add emitter power or open the aperture rather than raise gain.

Data rate is no obstacle: 327,680 pixels at 16 bits is 655,360 bytes per frame, and 60 fps is 39.3 MB/s — about 315 Mbit/s, inside one GigE link with packet overhead to spare. An apple crosses the 256 mm window in about half a second, so 60 fps delivers roughly 30 frames per fruit for track-and-grade; a photo-eye trigger gives one deterministic frame per cup instead.

Turning pixel counts into millimetres on the fruit is the same mapping described in imagesc axis coordinates: map pixels to real-world sizes.

Area scan or line scan

Singularised fruit — cups, rollers, singulators — belongs to area scan cameras: trigger one frame per fruit, hold exposure under a pixel of blur, grade, and hand the result to the ejector. Bulk streams on a wide belt belong to line scan: an encoder fixes the line rate, a SWIR line light covers the belt width, and every line is product. The pixel-speed-blur trade-offs are worked through in sorting conveyor camera selection: pixels, speed and blur; wide webs bring their own illumination problem, covered in line scan illumination uniformity.

The line rate comes from belt speed divided by pixel size: 500 mm/s ÷ 0.5 mm = 1000 Hz. Ejector timing, reject lanes and the dwell between detection and blow-off are their own engineering problem — optical fruit sorting engineering: vision, timing and rejection covers that side of the machine.

What each crop needs

Citrus drops to 970–1200 nm because the peel itself absorbs the water band strongly before light returns from the flesh; staying there costs you all depth. One honest line on nuts: SWIR does not measure aflatoxin, it removes the mold-damaged and insect-damaged nuts that carry the risk, and the regulated measurement stays a laboratory test on the retained stream.

Bench test before you order the line

1. Collect 40 fruits: 20 sound, 20 bruised by a controlled drop onto a hard plate.
2. Image both bands with the halogen-and-filter rig, at the geometry the lane will use.
3. Divide the frames and measure mean grey level in bruise versus sound tissue.
4. Require a contrast-to-noise ratio of at least 5 — the classic Rose criterion for reliable detection. Below that, change band or illumination before you change camera.
5. Cut every fruit open and confirm each dark region is damage, not shading or a surface wet patch.
6. Repeat with the belt running at line speed; contrast that survives the bench can still die in motion blur.

Food lines get hosed down. Keep the camera and lens behind a window rated for washdown, and check that the glass transmits at both of your bands before it is cut: choosing an infrared window material for a thermal sensor enclosure and a protective filter window for a vision lens cover the materials and the image cost.

When the water band is the wrong band

• Thick peel: citrus skin absorbs the water band before light returns from the flesh; use 970–1200 nm and fuse with colour.
• Dried product: raisins and figs hold little water, so the water band is weak; count on material contrast in the water-band/1650 nm ratio instead.
• Deep defects: internal browning centimetres inside an apple is invisible at every band in this range; that is transmission or density work, not reflection SWIR.
• Silicon cameras: a 1450 nm filter on a silicon sensor images nothing, because the sensor stops responding near 1000–1100 nm.

Before you order: the checklist

• InGaAs sensor, 900–1700 nm — silicon cannot see the water absorption peak.
• Lens rated for 900–1700 nm, image circle at least the sensor diagonal.
• Exposure at or under one pixel of blur at belt speed.
• Three pixels or more across the smallest defect you must catch.
• Two bands and a ratio image, not one band and a threshold.
• Diffuse illumination; enclosed, interlocked optical path.
• Washdown window that transmits at both bands.
• Bench contrast-to-noise of 5 or better on real bruised fruit.

Work the pixels in the camera resolution calculator, match the glass in the lens configurator, and for a multi-lane line, tell us the crop and the belt speed — we will size the rig with you.