Industrial machine vision has become a standard part of modern production. Cameras inspect labels, verify component placement, measure dimensions, read codes, and reject products that fall outside tolerance — at line speed, continuously, without human intervention. For a large class of inspection problems this works well, and has done for decades.
But there is a category of problem that conventional machine vision cannot solve, regardless of how good the camera is or how sophisticated the algorithm. When two materials look identical yet are chemically different, a colour camera has nothing to work with. A black polyester garment and a black cotton garment reflect visible light in much the same way. Two white powders can be entirely different compounds. Asbestos-containing cement board looks a great deal like ordinary cement board.
This article looks at what industrial machine vision does well, where its limits lie, and how hyperspectral imaging extends the same automation architecture — the same conveyor, the same trigger, the same programmable logic controller — into material identification.
What Is Industrial Machine Vision?
Industrial machine vision is the automated capture and analysis of images to make decisions inside a production process. A complete vision system consists of illumination, optics, an image sensor, a processing stage that turns pixels into a decision, and an interface that communicates that decision to the machinery around it.
The tasks it handles are familiar across manufacturing:
- Presence and absence checks — is the component fitted, is the seal in place
- Dimensional measurement — is the part within tolerance
- Code and character reading — barcodes, datamatrix, printed lot codes
- Surface defect detection — scratches, cracks, contamination visible on the surface
- Colour sorting — separating products by visible appearance
- Robot guidance — locating an object so a gripper can pick it
What defines an industrial vision system is not the camera on its own. It is whether the entire chain delivers a reliable, repeatable decision within the cycle time the line runs at. A camera that produces beautiful images but cannot keep up, or that gives different answers on Tuesday than it gave on Monday, is not a working industrial system. Repeatability is the currency.
Where Conventional Machine Vision Reaches Its Limit
A colour camera samples light in three broad channels — red, green, and blue. That is enough to reconstruct an image that looks correct to a human eye, and enough to measure geometry, contrast, and colour. It is not enough to determine what something is made of.
The limitation shows up in several recurring situations:
Materials that look alike but differ chemically. Polymer types in mixed plastic waste. Fibre types in mixed textiles. Mineral phases in crushed ore. In each case the visible appearance carries almost no information about composition.
Properties that are not expressed in visible colour at all. Sugar content in fruit. Moisture in a powder. Fat and protein distribution in a fillet. These are chemical quantities, and a three-channel camera cannot measure them.
Surfaces where colour masks the substrate. Dye and coating sit on top of the material. Once a garment is dyed black, or a board is painted, the visible surface tells you about the coating and nothing about what is underneath.
HySpex has made this point in the context of food inspection: machine vision techniques are being adopted rapidly across the industry, but because they are largely confined to RGB imaging, the range of problems they can address is correspondingly limited. The constraint is not the engineering quality of those systems. It is the amount of information three broad channels can carry.
Hyperspectral Imaging as an Extension of the Vision Stack
Hyperspectral imaging replaces the three broad channels with hundreds of narrow, contiguous ones. Every pixel in the resulting image carries a full spectrum rather than a colour value, and the shape of that spectrum — where light is absorbed, how strongly, across which wavelengths — reflects the molecular composition of the material at that point. Our overview of hyperspectral imaging covers the underlying principle in more depth.
For industrial use, the important point is architectural rather than conceptual. HySpex cameras use a pushbroom design, which images one line of the scene at a time and builds up the second spatial dimension through relative motion between camera and object. In a factory this is not a constraint but a convenience: in a typical industrial installation the camera is mounted statically above a conveyor and the belt motion supplies the scanning movement.
That means hyperspectral imaging does not require a new automation paradigm. The camera sits where a machine vision camera would sit. It is triggered the same way. It produces a decision that has to reach the same PLC within the same cycle time. What changes is the type of question the system can answer — from what shape and colour is this to what is this made of.
What an Industrial Inspection Platform Has to Deliver
An industrial inspection platform that is going to run in production has to satisfy four requirements at once. Nominal specifications on a datasheet are a starting point; these four determine whether the system actually works.
Repeatability under real conditions
Two optical distortions matter more than anything else for spectral repeatability: smile, where the wavelength calibration shifts across the field of view, and keystone, where a given spatial position falls on slightly different detector columns at different wavelengths.
Keystone is the more damaging of the two in an industrial setting. When it is significant, the spectrum recorded for one pixel is partly contaminated by the material sitting next to that pixel. The consequence is a system whose answers depend on what happens to be adjacent — which is another way of saying a system that is not repeatable. HySpex documents the relationship directly: as keystone increases, the relative error in the recorded spectrum rises proportionally, and at large keystone values the error becomes substantial enough to undermine any classification model built on the data.
HySpex Baldur industrial cameras keep keystone and smile below 15% of a native pixel and native band, and below 10% per effective pixel and effective band. Just as importantly, the correction is achieved in the optics rather than through a resampling step in software, so no additional processing stage is introduced between the detector and the data. The Baldur industrial design overview covers the optical reasoning in detail.
Model transfer between instruments
A property that matters enormously once a deployment grows beyond a single line: all Baldur cameras within the same wavelength range share the same centre wavelengths, and the spectral resolution is designed to sit at two bands specifically so that instruments remain spectrally comparable.
The practical effect is that a classification model developed on one camera can be used on another without recalibration or rebuilding. For an operator running four parallel lines, or replacing a unit after five years of service, this is the difference between a maintainable analytical infrastructure and a permanent modelling project.
Speed that scales with the problem
Industrial lines do not slow down to accommodate inspection. Baldur cameras address this in two ways. Acquisition speed scales directly with a reduction in the number of spectral channels read out — if an application only needs part of the spectrum, the camera runs faster in proportion. And all cameras operate in Integrate While Read mode, which allows almost the entire frame period to be used for exposing the detector. At high acquisition rates, where exposure time is the binding constraint, this matters considerably.
Light sensitivity is roughly four times higher in the Baldur line than in the HySpex Classic series, which is what makes short exposures at high belt speeds practical in the first place.
Calibration that can be audited
Baldur cameras are delivered with calibration traceable to NIST and PTB standards. For applications where an inspection result feeds a quality claim, a regulatory record, or a commercial specification, traceability is not a nicety — it is what makes the measurement defensible.
HySpex cameras also conform to the IEEE standard for hyperspectral imagers, which means a buyer can establish that a system will meet its target specifications before purchasing rather than discovering the gap during commissioning. In a market where headline specifications are not always comparable between vendors, a common standard is what makes a specification something you can hold a supplier to.
Worth noting for industrial buyers: because a close-up lens is required for industrial working distances, the cameras are calibrated together with the lens before delivery. The calibration therefore describes the optical chain that will actually be installed, not the camera in isolation.
Specifying a System: Effective Pixels and Minimum Object Size
One of the more useful things about HySpex's industrial documentation is that it states the dimensioning rule plainly rather than leaving it to the buyer to work out.
The starting point is effective pixel count, not native pixel count. Effective pixels are the total spatial pixels divided by the spatial resolution expressed in pixels — in other words, how many genuinely independent spatial samples the system delivers. If distortion per effective pixel exceeds 10%, the effective pixel count has to be reduced until it does not.
From there the specification follows a clear sequence:
- Define the smallest object that has to be detected.
- Allow at least two effective pixels across that object in one direction. One pixel is not enough; a single sample sitting partly on the object and partly on the background gives a mixed spectrum.
- Take the width of the conveyor belt. Together with step 2, this determines how many effective pixels the system needs across the belt.
- Work back to the native pixel count required to deliver that many effective pixels at the distortion limit.
A concrete illustration comes from HySpex's asbestos application note. Using a close-up lens with a 1 metre working distance and a 40° field of view, covering 70 cm across the scene, the system achieved a spatial resolution of 1.8 mm. Baldur cameras are offered with three lens configurations — 1 m working distance at 16° field of view, 1 m at 40°, and 1.9 m at 40° — which gives a practical range of belt widths and mounting heights to work within.
Effective pixel count is also the parameter that drives price, which makes price per effective pixel a more meaningful comparison between instruments than headline resolution figures.
Integration with Existing Automation
This is where a hyperspectral inspection system either fits into a plant or does not, and it is the part most often glossed over.
Triggering. All Baldur cameras can be triggered internally, and all support external triggering via TTL at several voltage levels as well as LVDS. That covers the signalling conventions found in most automation environments.
Model development and deployment. The software chain runs through Prediktera, a HySpex subsidiary. Breeze covers data acquisition, analysis, and application modelling in a single workflow, so a model can be developed and validated offline. Breeze Runtime then takes that model into production, performing real-time chemical quantification, classification, and object identification on material being scanned on-line. Models export from the offline package to the runtime engine, which runs on CPU or GPU.
Talking to the machinery. Connector software handles the part that automation engineers care about most: timing. 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 downstream. This software has been interfaced with several machine suppliers' systems to control PLCs in sorting applications, and HySpex's asbestos work demonstrates compatibility with third-party pneumatic and mechanical sorters and with robots.
Building it yourself. For integrators who prefer to own the software layer, every HySpex camera across the Baldur, Classic, and Mjolnir families ships with a high-end SDK and library for integration into third-party hardware and software. A NUC and framegrabber kit is offered as the standard processing option, and a third-party computer and framegrabber can be substituted.
Deploying a Vision System in a Production Environment
A scientific-grade instrument is only useful in a factory if it survives the factory. This is a legitimate objection to bringing research optics onto a production line, and it has a concrete answer.
The HySpex Modular Industrial Housing toolkit is an IP65-certified set of modules intended for exactly these conditions — installation over conveyor belts, on mining trucks, and inside material processing facilities, including extremely dusty ones. The modules can be combined in different configurations depending on the installation.
The single camera housing is dust-tight and water-resistant to IP65, rated from −20 °C to +45 °C, built in powder-coated steel and anodised aluminium, with the camera mounted over a fused silica window and an access door on the front panel. The detail that gives away practical experience is the window air nozzle: a stream of clean, dry compressed air across the window prevents dust building up, which reduces cleaning intervals in exactly the environments where cleaning is most awkward.
The rest of the chain is covered to the same standard. Sealed halogen lamps in rugged aluminium housings with fused silica windows, IP65 rated, with the same optional air nozzle. A PSU and DAU cabinet, also IP65, connected to the camera housing by flexible conduit. For the operator interface there is a choice between a handheld portable monitor, a permanently mounted IP65 industrial touchscreen, or a washdown-compatible workstation in stainless housing — the last of which is aimed squarely at food and fish processing. Systems can also be operated remotely.
For installations that are exposed rather than merely dirty, the HySpex Mjolnir IP65 configuration provides an environmentally protected camera on a pan and tilt head for unattended operation over extended periods.
There is also a design heritage argument worth making. The optical design underlying the Baldur series originates in airborne instruments and has been flying since 2003. It was built from the outset to hold calibration through vibration and mechanical stress, which is a reasonable proxy for what a production environment does to equipment.
Advanced Machine Vision in Practice
The case for advanced machine vision based on spectral data is stronger when it rests on deployments rather than principles.
Hazardous waste separation. HySpex's asbestos application note demonstrates classification of asbestos against concrete, ceramics, and terracotta in mixed demolition and renovation waste, using a Baldur SWIR camera over an industrial conveyor with a PLS-DA model built in Breeze. The reflectance spectra of the five material classes look broadly similar, yet each carries features distinctive enough to support a robust model — and all pieces were correctly assigned, including the smallest ones. The low keystone of the camera is what makes objects only a few pixels across identifiable at all.
Polymer identification. The mixed plastic waste study is the clearest available illustration of the RGB limitation. Five plastic types — PET bottle, PET sheet, PET-G, PVC, and polycarbonate — were imaged with a HySpex SWIR camera on a moving stage replicating conveyor motion. HySpex states the problem plainly: the five types look visually similar and cannot be reliably separated from a normal RGB image or by visual inspection, particularly since several appear transparent in the visible range. In the SWIR their spectra differ, though only slightly, because their chemical structures are closely related. That is precisely why spectral resolution matters: separating closely spaced absorption features is what makes the classification possible, and every piece in the test mixture was correctly assigned. Sorting polymers before re-melting is what determines the quality of recycled material and how many times it can be reused.
Measuring a property with no visible expression. The tomato quality grading deployment is the strongest illustration of what colour cannot tell you. A tomato farm producing 10,000 tonnes a year needed to grade its premium sweet tomatoes while they remained attached to the truss — a consumer preference that ruled out conventional single-fruit sorters. The difficulty was that deep red colour does not reliably indicate sweetness: damage to the plant stem during growth produces fully mature, deep red tomatoes that taste noticeably less sweet, and visual grading cannot detect the difference.
A Baldur V-1024 N mounted above the packing line now measures Brix value — sugar content — optically through the skin, in real time, without touching the fruit. The camera covers 400–1000 nm at 5.5 nm sampling, which reaches into the third overtone region of the NIR spectrum; combined with optimised illumination, that allows the measurement to reach into the fruit rather than reading only the surface. The system handles 40 trusses per minute, matching the pace of the two to four operators who cut and sort them, and presents colour-coded results on separate screens along the line as decision support rather than as automated rejection.
The commercial effect ran in both directions: fewer disappointing tomatoes reaching the premium line, and fewer genuinely premium tomatoes being downgraded and sold cheaply.
Commercialised food inspection. A Baldur V-1024 N sits at the centre of the Fish Quality Analyzer, developed in the FHF-funded KVASS project alongside Maritech, NOFIMA, Lillebakk, Lerøy, Havfisk, and Prediktera, and now offered commercially by Maritech. This is a completed path from research project to product on the market.
Textile fibre separation. NEO has worked with Norsk Tekstilgjenvinning and Steco Miljø on a fully integrated optical textile sorter capable of classifying between four and over a hundred fibre types and mixtures.
Choosing Between Conventional and Hyperspectral Vision Systems
The decision is more straightforward than the technology suggests, and it comes down to a single question: is the property you need to measure expressed in visible colour and geometry?
If it is — dimensions, position, printed codes, visible surface defects, colour differences a person could see — conventional machine vision is the right answer. It is cheaper, faster, simpler to integrate, and supported by a mature ecosystem. Choosing hyperspectral for these tasks means paying for information you do not need.
If it is not — composition, chemistry, concentration, material type, internal quality — then no amount of resolution or processing sophistication will extract it from three broad channels. That is the boundary where hyperspectral imaging becomes not merely better but necessary.
In practice many production lines end up running both, with conventional vision handling geometry and presence checks while a spectral system handles material identification and quality measurement. They answer different questions, and the systems coexist comfortably on the same line. Our overview of hyperspectral imaging systems covers how the complete system is configured, and the Industrial turnkey solution page sets out how HySpex delivers this as an integrated package.
FAQ – Industrial Machine Vision
What is industrial machine vision?
Industrial machine vision is the automated capture and analysis of images to make decisions inside a production process. A complete system includes illumination, optics, an image sensor, processing that converts pixels into a decision, and an interface to the surrounding machinery. Typical applications include dimensional measurement, presence and absence checking, code reading, surface defect detection, and colour sorting.
How is hyperspectral imaging different from conventional machine vision?
A conventional camera records three broad colour channels, which is sufficient for geometry, contrast, and visible appearance. A hyperspectral camera records hundreds of narrow, contiguous spectral bands for every pixel, which reveals molecular composition. The practical consequence is that hyperspectral systems can distinguish materials that look identical and can measure chemical properties that have no visible expression at all.
Can a hyperspectral camera work with an existing PLC-based sorter?
Yes. HySpex Baldur cameras support external triggering via TTL at several levels and LVDS, and connector software handles frame and object timing so that a decision reaches the control system at the right moment. This software has been interfaced with several machine suppliers' systems to control PLCs in sorting applications, and the cameras have been used with third-party pneumatic and mechanical sorters as well as robots.
How fast can an industrial hyperspectral vision system run?
Speed depends on how much of the spectrum the application needs. Acquisition rate scales directly with reductions in the number of spectral channels read out, so an application requiring only part of the spectral range runs proportionally faster. Integrate While Read mode allows nearly the full frame period to be used for exposure, and the roughly fourfold light sensitivity advantage of the Baldur line over the Classic series makes short exposures practical at production speeds.
What protects an industrial inspection platform in a dusty or wet environment?
The HySpex Modular Industrial Housing toolkit provides IP65-rated enclosures for the camera, sealed lamps, and the power and data acquisition electronics, rated from −20 °C to +45 °C. An air nozzle across the camera window prevents dust accumulation. Operator interface options include an IP65 mounted touchscreen and a washdown-compatible stainless workstation for food processing environments.
How do I work out what spatial resolution I need?
Define the smallest object that has to be detected, then allow at least two effective pixels across it in one direction. The width of the conveyor belt then determines the total effective pixel count required, from which the necessary native pixel count follows. Effective pixel count — rather than headline resolution — is the parameter that determines both capability and price.
Discuss Your Industrial Inspection Requirements
At HySpex we develop hyperspectral imaging systems for industrial inspection, sorting, and process monitoring, delivered as integrated packages combining Baldur cameras, Prediktera software, industrial housing, and the integration support needed to connect them to existing automation. If your application involves material identification, quality measurement, or sorting where conventional machine vision has reached its limit, a technical discussion about your specific line, throughput, and object size is usually the most productive starting point.
