What the line actually needs
German collectors handle around 1.3 million tonnes of textiles a year and sort them into as many as 350 fractions. Sorting costs run to roughly 0.68 euro per kilogram. At those volumes and margins the economics are unforgiving: a decision is made once, in under a second, and must be good enough that the fraction it produces holds its value.
That is not the same target as a laboratory result. A laboratory answers the question "what is this material made of, to a stated tolerance". A sorting line answers "which of 350 destinations does this item belong in, and can I defend that later". Chasing the first target with the second budget is why so many automation projects in this industry stall.
Laboratory methods, and why they stay in the laboratory
The established methods are burn tests, quantitative chemical analysis by dissolution, optical microscopy and spectroscopy. They work. They are the reference against which anything else is judged, and for a disputed claim they remain the right answer.
- They are destructive, or near enough that the sample does not go back into the stream.
- They are priced per sample, which makes testing at scale an arithmetic that never closes.
- They take hours to days, against a line that moves items in milliseconds.
- They read the material, so they say nothing about where the item came from or how it must be treated.
The last point is the one that gets overlooked. Even a perfect laboratory result tells a sorter what the fibre is. It does not tell them the brand, the origin, the care requirements or whether anyone ever certified anything about it. Those are the fields that decide whether an item is worth reuse, resale or shredding.
Sensor based sorting, fairly described
Near infrared and hyperspectral sorting lines are the real industrial answer to material identification, and they deserve to be described accurately rather than dismissed. They read a spectrum reflected off the fabric and infer the polymer, at speed, without touching the item. For bulk separation of a mono material stream they are the correct tool and there is no software substitute for them.
Their limits are documented and are limits of physics rather than of engineering effort.
- They read the outer surface of the fabric, so a lining, a coating or a laminate is what gets measured.
- Blends return a mixed signal that is hard to resolve into a stated ratio.
- Dark dyes and prints absorb the signal the measurement depends on.
- Multiple layers, trims and seams present a different material from the one that dominates the garment.
- Most importantly: a spectrum carries no provenance, no brand, no care instruction and no certification. It identifies material, never the item's history.
The data source nobody reads
Nearly every garment in a collection stream already carries the answer, sewn into a seam. The composition and care label is not an informal note. It is a regulated document. Regulation (EU) 1007/2011 fixes the fibre names that may appear on it and requires the composition to be stated. ISO 3758 fixes the care symbols. The information a line needs is physically present on the item, in a controlled vocabulary, placed there by law.
And it is thrown away. Not because it is worthless, but because reading it has always been a human task: a label in one of two dozen languages, faded, folded, cut, upside down, sometimes three labels disagreeing with each other. At 350 fractions and a fraction of a euro per kilogram, nobody can afford a person to read every one.
Reading it at line speed
A capture station photographs each garment from several angles as it moves. The composition, care, origin and brand fields are read from whatever labels are present, in whatever language, and each field is scored. Values are normalized into controlled EU vocabularies, so a label reading 62% CO and a label reading 62% Baumwolle produce the same field. The result is an item record, and every value in it keeps a link back to the image region and the document it was read from.
Two properties make this usable rather than merely clever. The first is throughput: 23,800 items an hour across 16 concurrent capture streams, with routing decisions inside a 50 ms ceiling, which is what it takes to sit on a line rather than beside it. The second is the confidence gate. Roughly a quarter of items clear a 0.9 threshold and route automatically. The rest go to assisted review, where a person sees the item, the label crop and what the system read, and decides.
That second number is the honest one. A system that claimed to route everything automatically would be a system that guesses on the hard items, and the hard items are exactly where a wrong fraction destroys value. The point is not to remove the human. It is to spend the human's time only where it changes the outcome, and to know in advance which items those are.
Complementary, not substitutive
The two approaches answer different halves of the same question, and the interesting case is where they disagree. A spectrum reading polyester against a label declaring 100% cotton is not an error to be resolved silently. It is a finding: a coating, a mislabelled item, a lining dominating the surface, or a garment whose label no longer describes it. A record that carries both readings and their disagreement is worth more than either reading alone, because the disagreement is information.
| Approach | What it reads | What it cannot read | Speed per item |
|---|---|---|---|
| Laboratory analysis | Material composition to a stated tolerance. | Provenance, brand, care, certification. Whether the fibre was previously used. | Hours to days, per sample. |
| NIR and hyperspectral sensing | Surface polymer identity, without contact. | Anything below the surface layer. Provenance, brand, care, certification. | Milliseconds, continuous. |
| Label reading with confidence | Composition, care, origin and brand as declared, each scored and traceable. | What the material is when the label is absent, illegible or wrong. | Inside a 50 ms routing ceiling. |
None of the three is the whole answer. But only one of them produces a record an auditor can read, and it is the one built from the document the garment was already carrying.