The recycling scanner that reads a garment beyond its label

Turn a hoodie inside out and a fibre scanner may give you a different answer. Matoha uses just such an example in its FabriTell documentation: a garment labelled 65% cotton and 35% polyester can have a predominantly cotton outer surface and a fluffy inside that is mostly polyester. A surface measurement and a whole-garment label are describing different things.[1]

That distinction sits at the centre of an intriguing industrial use for AI. Sorting discarded clothing by what it is made from could provide recyclers with a more dependable raw material. The hard part is translating a reading from the fabric into a decision about where the entire item can go.[5]

A chemical clue in reflected light

Near-infrared spectroscopy measures how a material responds to wavelengths beyond visible red light. Because that response reflects chemical bonds, it can help distinguish fibres whose appearance alone is ambiguous. NIST describes a typical sorting system as a sensor supplying measurements to a classification model: software, often using machine learning, associates the measured spectrum with a fibre category.[2]

The roles are complementary. The instrument supplies evidence from the material. The trained model interprets patterns in that evidence. A system that identifies fabric composition this way is doing something different from reading the printed care label, so the label need not survive for the measurement to be useful.

PICVISA says its ECOSORT TEXTILE installation at ESSAIMONS in France combines ordinary colour vision, near-infrared sensing and AI to identify composition in mixed textile streams. The manufacturer describes the intended output as more homogeneous material for recycling. That is a deployment account from the supplier; it does not establish an independently measured recovery rate for the plant.[3]

One reading cannot describe every garment

Matoha’s published limits make the engineering problem concrete. Its device can miss components present below 5%, or fibres hidden within the weave. It cannot read samples containing carbon black, which absorbs the infrared light. And three-component blends may be identified mainly through their two dominant constituents. These are the supplier’s stated limits for its equipment, not performance specifications for every textile scanner.[1]

The hoodie example suggests a practical question for sorting-line designers: does the system measure enough of the item to support the decision being made? A reading of the sleeve, lining or outer panel may each be useful, but combining them requires a rule about which components the receiving process can tolerate. Confidence in a local measurement and confidence in the routing decision are separate judgements.

Teach the model what the fabric really is

Training data are another part of that chain. NIST’s NIR-SORT provides characterised fabrics and spectra with documented provenance. It addresses a problem the institute identifies in the field: many reference libraries are proprietary or assembled privately, making results difficult to compare. The resource includes measurements from different instruments alongside fabric imagery.[2]

A complementary research resource is OpenTextile-NIR, described by VTT researchers in the April 2026 issue of Data in Brief. It contains hyperspectral and colour imagery from 71 textile samples, with composition information supplied by manufacturers. Its millions of spectra are measurements within those samples, rather than millions of independent garments. The distinction matters when deciding how much variety a model has actually encountered.[4]

For a builder, these resources make a more informative development question possible: which fibres, blends and appearances are represented, and which are still unfamiliar? Adding many measurements from the same piece of fabric cannot, by itself, answer how a system will handle a different garment.

The destination determines the useful sort

Fashion for Good’s Feedstock Activation Europe project starts with the requirements of textile-to-textile recyclers: composition, colour and contaminants, at a usable price and volume. It distinguishes sorting from preparation such as cutting, shredding, removing trims or separating blends. Its proposed regional hubs would combine those operations and assemble streams tailored to different recyclers. The commercial benefits remain a project objective.[5]

That gives the AI task a sharper purpose. Calling something “mostly cotton” is only an intermediate result. The valuable decision is whether that material, after the necessary preparation, meets a particular buyer’s specification. A more detailed classifier helps when its distinctions correspond to a route the material can actually take.

The scanner earns its place by helping an old garment find its next industrial use. Reading the fibres is the beginning of that journey.

Report an error or suggest a correction