Textile machine vision faces a problem that most machine vision applications do not. In a typical inspection task the property being measured is visible — a dimension, a position, a printed code, a surface flaw. In textiles, the property that matters most is invisible. What a garment is made of leaves almost no trace in its visible appearance.
This is not a resolution problem, and it cannot be solved with a better colour camera. Two t-shirts dyed the same shade of black, one cotton and one polyester, reflect visible light in essentially the same way. The information that separates them sits outside the visible spectrum entirely.
This article looks at how textile machine vision identifies fibre composition, what the software layer has to do to turn that identification into a usable signal, and how the whole chain connects to the machinery that acts on the result.
What Is Textile Machine Vision?
Textile machine vision is the use of automated imaging to inspect, classify, or identify textile material in a production or processing line. The applications divide into two groups with quite different requirements.
Appearance-based tasks are conventional machine vision applied to fabric: detecting weaving faults, print registration errors, stains, holes, shade variation between batches, or foreign objects. Here a high-resolution colour or monochrome camera is the right tool, because the defect is genuinely visible.
Composition-based tasks are the harder category: determining fibre type, detecting blends, identifying contaminants that look like textile, or separating a mixed stream into material fractions. These require spectral information rather than colour information, and they are the reason hyperspectral imaging has become central to textile processing.
The distinction matters when specifying a system, but it is not an either-or choice. A deployed textile sorting line typically does both at once. The automated sorter developed by Norsk Elektro Optikk with Norsk Tekstilsortering and Steco Miljø identifies fibre content alongside colour, brand logos, and accessories such as buttons, zips, and prints — spectral identification and appearance-based recognition working on the same items. An operator who needs both usually needs both sensing approaches on the line, not one camera attempting everything.
Why RGB Vision Cannot Determine Fibre Content
A colour camera samples light in three broad bands. That is sufficient to reproduce an image a human would recognise, and to measure anything expressed as colour, contrast, or geometry. Fibre composition is expressed as none of those things.
Three factors compound the problem in textiles specifically.
Dye dominates the visible signal. Colour comes from the dye applied to the fibre, not from the fibre itself. The same dye can be applied to cotton, polyester, viscose, or a blend, producing garments that look identical and are made of entirely different materials.
Construction confuses appearance. Knit versus woven, yarn thickness, surface finish, and brushing all change how a fabric looks without indicating what it is. A brushed polyester fleece can read visually as something quite different from a smooth polyester shirt, despite being the same polymer.
Labels are unreliable at volume. Care labels would answer the question in principle. In practice, at processing volumes, a large share are faded beyond reading, cut out, detached, printed in an unexpected language, or simply wrong. No system that depends on labels can be relied on for material recovery.
HySpex has documented the same underlying limitation in an adjacent material stream. In the mixed plastic waste study, five polymer types were shown to be visually similar and impossible to separate accurately from a normal RGB image or by visual inspection — several appear transparent in the visible range. The polymers were separable in the shortwave infrared, but their spectral differences were small, because their chemical structures are closely related. Textile fibres present the same structural challenge.
How Spectral Identification Works on Textile Fibres
Hyperspectral imaging replaces three broad colour channels with hundreds of narrow, contiguous spectral bands, recording a full spectrum for every pixel in the image. The shape of that spectrum reflects molecular composition, because different chemical bonds absorb light at characteristic wavelengths.
For textiles, the informative region is the shortwave infrared, roughly 1000 to 2500 nm. This is where overtone and combination absorptions from the molecular bonds that make up fibre polymers appear, and it is why SWIR cameras rather than visible-range cameras are used for fibre identification.
The broad chemical groupings behave differently:
Cellulosic fibres — cotton, viscose, lyocell, linen — are built on cellulose, and their SWIR spectra are shaped by hydroxyl and carbon-hydrogen absorptions characteristic of that structure. Distinguishing cotton from regenerated cellulosics such as viscose is harder than distinguishing either from a synthetic, because the underlying polymer is chemically similar.
Protein fibres — wool, silk — contain amide linkages and nitrogen-hydrogen bonds absent from both cellulosics and most synthetics, which gives them relatively distinctive spectral features.
Synthetic polymers — polyester, polyamide, acrylic, elastane — each have absorption features arising from their specific chemistry. Polyester, as a polyester of an aromatic acid, carries features that separate it clearly from cellulose. Distinguishing between different synthetics can be more demanding, for the same reason the plastic study noted: closely related chemical structures produce closely spaced spectral features.
Crucially, these features come from the fibre polymer itself rather than from the dye. That is what makes spectral identification robust across colour — a black polyester garment and a white polyester garment carry recognisably the same SWIR signature, which is precisely what RGB imaging cannot achieve.
The practical requirement that follows is spectral resolution. When the differences between classes are small, an instrument that blurs adjacent spectral features together loses the information that separates them. This is the same conclusion HySpex drew from the plastics work: high spectral resolution is what makes separation of chemically similar materials possible.
The Hard Cases
A textile vision system that handles clean single-fibre garments and fails on everything else is of limited use, because a real intake stream is not clean.
Blends are the central difficulty. Cotton-polyester in varying proportions, wool blends, cellulosics mixed with synthetics, and elastane present at a few percent across a wide range of garments. A blend does not produce a new spectrum so much as a mixture of contributing spectra, weighted by proportion. Detecting that an item is a blend is one problem; estimating the ratio is a harder one, and often the commercially decisive one — chemical recycling routes are considerably less tolerant of a second fibre than mechanical routes are.
Coatings and finishes sit on the surface and can dominate the measured spectrum. Waterproof membranes, prints, and heavy surface treatments may mask the substrate rather than merely colouring it.
Layered and multi-component garments present several materials in one item — a lined jacket, a garment with elasticated panels, a shirt with a different collar fabric. A single classification per item is not always meaningful; spatially resolved classification, where different regions of the same garment are classified separately, is often what the application actually requires. This is where imaging spectroscopy differs usefully from point spectroscopy: it gives a classification per pixel rather than per item.
Irregular presentation. Garments arrive folded, overlapping, crumpled, and at varying heights above the belt. Unlike rigid plastic flakes or fruit on a conveyor, textiles do not present a flat, consistent surface, which affects both illumination geometry and the reliability of a per-item decision.
Textile Vision Software: From Model Development to Runtime
Hardware produces spectra. Textile vision software is what turns spectra into decisions, and it is generally where a project succeeds or stalls.
The workflow through Prediktera, a HySpex subsidiary, runs in two stages.
Model development happens in Breeze, which covers data acquisition, analysis, and application modelling in a single environment. For a textile application this means scanning representative garments of known composition, building a spectral library for the fibre classes the operation needs to separate, and training and validating a classification model against held-out samples. HySpex's application notes in adjacent materials illustrate the pattern: in the asbestos work, a classification model was trained on one sample set and tested against four others, which is the kind of validation discipline a textile model needs before it goes near a production line.
Preprocessing matters more than it might appear. Baseline correction, smoothing, and normalisation reduce the influence of illumination variation and surface geometry — both significant with irregularly presented garments — before any classification is attempted. Wavelength or region selection can then reduce dimensionality, which both improves model stability and increases achievable speed.
Deployment moves the validated model to Breeze Runtime, which performs real-time classification, object identification, and chemical quantification on material being scanned on-line in process. Models export from the development environment to the runtime engine, which can use CPU or GPU processing.
A detail worth knowing for anyone building a multi-line operation: Breeze Runtime supports cameras from several manufacturers as well as HySpex Baldur, Classic, and Mjolnir instruments, with a hardware SDK available for cameras outside the supported list. For an integrator, this reduces the risk of the software layer becoming a single-vendor commitment.
Connecting Vision Software to Sorting Machinery
Identification that never reaches the machinery sorts nothing. This is the least glamorous part of a textile vision system and the part most likely to be underestimated.
The timing problem is specific. A garment is imaged at one point on the belt and has to be acted on at another, some distance and some milliseconds later. The system needs to know not just what the item is but which item it was and when it will arrive at the ejection point. HySpex's connector software addresses exactly this: it tracks the mid-exposure moment of each frame — or the position and time of individual objects — and determines when a decision has to be issued for that object to be acted on correctly downstream.
On the output side, this software has been interfaced with several machine suppliers' systems to control PLCs in sorting applications. For textiles specifically, HySpex describes its role in the Norwegian automated sorter as supplying a complete hardware and software solution for real-time classification, interfacing with the machine's PLC to control the valves that perform the physical separation. The asbestos application note confirms the same compatibility with third-party pneumatic and mechanical sorting machines and with robots — the same class of actuator used in textile lines.
The textile sorter built for Norsk Tekstilsortering illustrates how the layers combine in practice: real-time classification hardware and software from HySpex, conveyor and sorting hardware from Steco Miljø, and an operator — a subsidiary of recycling company Norsk Tekstilgjenvinning — running the plant alongside its own recycling facility in Sandefjord. The system classifies between four and more than a hundred fibre types and mixtures. Nobody supplies all of that alone.
Triggering is handled through conventional automation signalling. HySpex Baldur cameras can be triggered internally and support external triggering via TTL at several voltage levels and LVDS, so synchronisation with existing line control does not require custom electronics.
Building on the SDK: Options for Integrators
Not every integrator wants to adopt a vendor's analytical software. For those building their own processing chain, every HySpex camera across the Baldur, Classic, and Mjolnir families ships with a high-end SDK and library for integration into third-party software and hardware, and HySpex describes itself as compatible with the third-party processing software suppliers on the market.
This matters commercially as much as technically. A machine builder with an established software platform and existing customer relationships can treat the camera as a component rather than adopting an entire ecosystem. HySpex supplies camera modules as well as fully integrated systems, and a platform-independent SDK is available for teams managing the software layer themselves.
For processing hardware, a NUC and framegrabber kit is offered as the standard option, with third-party computer and framegrabber combinations supported as an alternative.
What to Look for in a Textile Vision System
A few specification questions separate systems that work in a textile line from systems that work in a demonstration.
Spectral range and resolution. Fibre identification is SWIR work. Resolution has to be sufficient to separate chemically similar classes — a blurrier instrument delivers less information per unit of data, and the classes that matter in textiles are frequently close together.
Spatial resolution against the smallest thing you need to identify. HySpex states the rule directly: define the smallest object to be detected and allow at least two effective pixels across it in one direction. Belt width then determines total effective pixel count. For textiles this applies to component parts as much as whole garments — a small elasticated panel is a different requirement from a whole jacket.
Optical distortion. Keystone and smile determine whether a spectrum belongs to the pixel it was recorded from or is contaminated by its neighbour. Contaminated spectra produce models whose answers depend on what happens to be adjacent, which is the definition of an unrepeatable system. HySpex Baldur cameras hold keystone and smile below 15% per native pixel and band, and below 10% per effective pixel and band, corrected in the optics rather than by resampling.
Model transferability. All Baldur cameras within the same wavelength range share the same centre wavelengths, and spectral resolution is designed at two bands specifically for cross-instrument consistency. A model developed on one camera can be used on another without rebuilding — which for a multi-line operation, or one that will replace units over time, is the difference between an asset and an ongoing project.
Illumination. Spectral measurement depends on the light source as much as the camera. Broadband halogen illumination is used across HySpex's industrial application notes, and for protected installations sealed IP65 lamp assemblies are available as part of the modular industrial housing toolkit.
Environmental protection. Textile processing generates fibre dust in quantity. IP65 enclosures with an air nozzle across the camera window are designed for precisely this, keeping the optical path clear without frequent manual cleaning.
Our overview of industrial machine vision covers the integration and specification principles across industries, and hyperspectral software gives a broader view of the software landscape. For the regulatory and market context driving textile processing capacity, see our article on textile recycling.
FAQ – Textile Machine Vision
What is textile machine vision?
Textile machine vision is the use of automated imaging to inspect, classify, or identify textile material in a production or processing line. It covers appearance-based tasks such as detecting weaving faults or stains, where a conventional camera is appropriate, and composition-based tasks such as identifying fibre type or detecting blends, which require spectral rather than colour information.
Why can't a normal camera identify fibre type?
Fibre composition is not visible. Colour comes from dye rather than from the fibre, so the same shade can be applied to cotton, polyester, viscose, or a blend, producing visually identical garments made of different materials. Fabric construction and surface finish further change appearance without indicating composition. The features that distinguish fibres lie in the shortwave infrared, outside the range a colour camera records.
Can dyed and printed garments still be identified?
Yes. The spectral features used for identification arise from the fibre polymer itself rather than from the dye, so colour does not obscure them. Heavy surface coatings and membranes are a different matter, since these can mask the substrate rather than merely colouring it.
Can a vision system detect fibre blends?
Detecting that an item is a blend is achievable, since the measured spectrum reflects a weighted mixture of the contributing fibres. Estimating the proportion is harder, and how much precision is required depends on the downstream route — chemical recycling is considerably less tolerant of a second fibre than mechanical recycling.
What does textile vision software actually do?
It performs two distinct jobs. In development, it supports scanning reference material, building spectral libraries, preprocessing data, and training and validating classification models. In production, it runs the validated model in real time on incoming data and issues a decision for each item, with timing logic that ensures the decision reaches the sorting mechanism when that item arrives at it.
Can it be connected to an existing sorting line?
Yes. HySpex cameras support external triggering via TTL and LVDS, and connector software handles frame and object timing so that decisions reach the control system at the correct moment. This software has been interfaced with several machine suppliers' systems to control PLCs, and the cameras have been used with third-party pneumatic and mechanical sorters and with robots. For integrators building their own software layer, an SDK is supplied with every camera.
Discuss Your Textile Identification Requirements
At HySpex we develop hyperspectral imaging systems used to identify materials that conventional vision cannot distinguish, supplied either as camera modules for integration into your own platform or as complete imaging systems with real-time processing software. A platform-independent SDK is available for teams that prefer to own the software layer.
If your project involves fibre identification, blend detection, or integrating spectral vision into an existing line, a technical discussion about your fibre classes, throughput, and control interface is usually the most productive place to begin.
