Field evidence

Real frames. Real detections. Measured in millimetres.

Everything below was cut from full-width frames stored by the inspection software on a deployed production line. The outline is the region the engine reported; the size is what it measured. Hover any crop to see the raw fabric without the outline.

107
production lots, one line
14,871
full-width frames stored with position
2,400+
regions confirmed, rejected or retyped by operators
22
defect categories in that customer's own scheme

Read from the line's inspection database. Lot names, customer and fabric article numbers are withheld.

Fabric range on that line

Eleven grounds from one line — one engine, one configuration.

Olive green fabricRed fabricCamel fabricIndigo fabricSky blue fabricNavy fabricSand fabricPowder blue fabricSage fabricSteel grey fabricCharcoal fabric

Crops from the centre of one frame per lot, colour as captured after flat-field correction. Dark, saturated and light grounds all run on the same engine with the same configuration; each fabric's reference is built automatically.

crop_free

The outline

Is the polygon the engine stored for the region, drawn 1:1 on the crop. Not a bounding box, not redrawn by hand.

straighten

The size

Width × height in mm from the calibrated pixel scale (0.137 mm/px on this line). "Field" is the width of the crop shown, for scale.

fact_check

The selection

We chose regions across fabrics and sizes, favouring operator-confirmed ones. We did not choose the easy ones only; several are near the limit of what a person can see on screen.

By defect type

Twenty-nine real detections, shown by category.

These are some of the categories in this line's scheme. Categories are defined and named by the customer from their own defect images — as many as they need, not a fixed ten.

Scale

A 0.21 mm spot, somewhere in 1.5 metres of width.

This is the problem in two pictures. First, what one camera sees across the fabric, with the reported region boxed. Then the same region at sensor resolution, where a person can finally see it.

Full camera view of green fabric across the fabric width, with a small marked box left of centre
One camera's view across the fabric. The box marks the reported region. The dark band on the right is the machine backing beyond the selvedge — the engine trims it before inspecting.
The same colour spot at sensor resolution, outlined by the engine, 21 by 27 mm

The same region, at sensor resolution.

Colour spot, 21 × 27 mm, outlined as the engine reported it. The window shown is about 125 mm across — roughly one twelfth of the fabric width above. At production speed nobody is looking at this square; the engine looks at every one of them.

What this page does not claim

Three honest notes.

False alarms exist. Lighting flicker, edge glow and marks the mill considers acceptable produce regions that the operator rejects. The feedback loop is there because of them, and the reject rate is one of the metrics we track per machine.
Detection rate above 95% is what deployed lines measure against operator review; the exact figure moves with fabric, speed and lighting. We measure it with the same tooling on every machine, so a machine builder can compare machines and versions.
This is one line. It is a real one, running production, and the figures above are read from its database. It is not a lab bench and not a marketing sample set.

See it on your own fabric.

Bring the fabric that gives you the most trouble to a pilot machine. We run it and hand back the record — every region, its size, its position and the image it came from.