How to Design a SWIR Material Classification System: Wavelength Selection, Lens FOV, Reference Samples, Calibration and Decision Thresholds

Designing a reliable SWIR material classification system requires much more than selecting a camera, pointing it at a sample and training software to recognize different brightness levels. In industrial environments, material identification succeeds only when the complete optical and decision chain is engineered together: the wavelength range must expose a repeatable physical difference between classes, the SWIR camera lens must provide the required field of view without sacrificing critical spatial sampling, representative reference samples must define real production variability, calibration must separate material behaviour from illumination and system drift, and decision thresholds must provide enough margin to distinguish accepted, rejected and unknown material populations. If any one of these stages is weak, classification performance can deteriorate even when the initial laboratory images appear impressive.

A strong design process therefore begins with the question “What material difference must the machine prove?” rather than “Which lens should I buy?” The intended decision may be polymer A versus polymer B, correct raw material versus wrong material, acceptable product versus contamination, dry versus excessive-moisture condition, approved coating versus incorrect coating, or one material composition versus another. Once the classes are defined, the engineer can determine whether they separate usefully across approximately 900–1700 nm and then design the optical field around the smallest region that must be classified.

The dedicated Kyptec Automation® SWIR Camera Lens collection provides 8.5 mm, 12.5 mm, 25 mm, 35 mm and 50 mm focal lengths for 900–1700 nm imaging. Current product pages specify 2 MP resolution, 2/3-inch sensor format, F1.4 aperture and C-Mount, with material identification among the intended industrial applications. This gives OEMs a practical focal-length range for translating verified spectral separation into a production-ready inspection geometry.

Start With the Classification Question Before Choosing Wavelength or Lens

A material-classification project should begin with a precisely defined production decision. “Identify materials with SWIR” is too broad because the required optical architecture changes depending on whether the machine must distinguish two known materials, separate ten known classes, detect an unknown material, identify a small contaminant inside a host product or determine whether one incoming raw material matches an approved specification.

A useful specification defines the classes and the consequence of each decision. For example, the machine may need to classify Material A, Material B and Material C while sending anything else to an unknown category. Another system may need only two decisions: approved material or abnormal material. A third may need to distinguish a 90/10 blend from an 80/20 blend, which is a much harder problem because the populations are compositionally close.

The narrower and more measurable the decision, the stronger the wavelength selection and threshold design can become.

Step 1: Collect Representative Material Samples Before Optical Optimization

The reference samples determine what the classifier ultimately learns. If the development set does not represent real production variation, the machine can appear accurate during commissioning and fail when raw materials, suppliers or environmental conditions change.

Each material class should therefore include several independently verified samples covering realistic variability such as supplier, production batch, colour, thickness, surface finish, moisture level, temperature, recycled content, additive concentration or other relevant differences.

The goal is not to capture one perfect spectral signature per material. It is to establish a spectral population representing what the material is allowed to look like in production.

This principle is particularly important for buyer-intent applications such as automated material identification, polymer sorting, raw-material verification, recycling inspection and inline quality control.

Step 2: Find Wavelengths That Separate Populations, Not Individual Samples

A useful SWIR wavelength is not simply one where two selected samples look different. It must preserve separation between the complete distributions of the material classes.

Suppose Material A measures between 0.52 and 0.58 normalized reflectance at one wavelength while Material B measures between 0.70 and 0.77. The gap between the populations is useful.

At another wavelength, Material A may vary from 0.40 to 0.62 and Material B from 0.55 to 0.73. Their average values differ, but the populations overlap substantially.

The first wavelength can therefore provide stronger production classification even if the second produces a larger difference between two hand-selected specimens.

A practical class-separation metric can be expressed as:

Separation = |μA − μB| / √(σA² + σB²)

where (μ) represents mean material response and (σ) represents within-class variation.

The design objective is to maximize separation relative to normal variability, not absolute brightness difference.

Step 3: Determine Whether One Wavelength Is Enough

Some industrial material-identification problems can be solved with one carefully selected wavelength. Others require two or more bands because the materials overlap at one wavelength but separate when their spectral shapes are compared.

A two-band normalized relationship can be written as:

N = (I₁ − I₂) / (I₁ + I₂)

where (I₁) and (I₂) are measured intensities at two selected wavelengths.

Band ratios or normalized differences can be valuable because they represent relative spectral behaviour rather than relying completely on absolute intensity.

However, more wavelengths do not automatically mean a better production system. Every additional band increases illumination, calibration, acquisition, processing and timing complexity. The optimum design uses the minimum number of wavelengths that provides stable separation across all required classes and production conditions.

Step 4: Identify the Hardest Material Pair

If a system must classify six materials, the two easiest-to-separate classes should not define the optical design.

The hardest pair determines the required system performance.

For each material combination, calculate or experimentally evaluate how much the spectral distributions overlap. The pair with the smallest decision margin deserves the greatest engineering attention.

For example, Materials A, B, C and D may separate strongly, while Materials E and F are almost identical except around one narrow part of the SWIR range. That spectral region may become decisive for the complete machine.

A strong OEM workflow therefore builds a class-separation matrix before lens selection.

Step 5: Decide Whether the System Needs Exact Identification or Abnormal-Material Detection

Exact material classification and anomaly detection are different objectives.

An exact classifier may output:

Material A
Material B
Material C
Material D

An abnormality-based system instead asks whether the measured material belongs inside the approved population.

For incoming quality control or contamination detection, the second approach can sometimes be safer because it does not force an unfamiliar material into one known category.

The production logic may become:

approved known material → accept

known wrong material → reject

outside all approved populations → unknown/reject

This unknown state is especially important when unexpected materials can enter the line.

Step 6: Define the Physical Region That Must Be Classified

The system may identify:

an entire object;

one region on an object;

a small contamination area;

a thin coating;

a continuous sheet;

or a granular material stream.

This distinction directly influences lens FOV.

Whole-object classification can often use an average response from hundreds or thousands of pixels. A 2 mm foreign fragment requires much finer spatial sampling because its signature can be diluted by surrounding material.

Before selecting focal length, define the smallest classification region that must remain independently measurable.

Step 7: Calculate Object-Side Pixel Sampling

Suppose the image width is 1600 pixels and the required horizontal FOV is 400 mm.

Object-side sampling is:

400 mm ÷ 1600 = 0.25 mm/pixel

A 5 mm material region spans approximately:

5 ÷ 0.25 = 20 pixels

A 1 mm region spans approximately four pixels.

If the FOV increases to 800 mm:

800 ÷ 1600 = 0.50 mm/pixel

The same 1 mm material region now spans approximately two pixels.

This demonstrates why field of view should not be maximized automatically. Wide coverage is commercially attractive, but if the smallest material region becomes under-sampled, classification confidence can fall.

Step 8: Select Lens FOV From the Material Decision

The correct SWIR camera lens is the one that covers the required production region while keeping enough pixels on the minimum classification target.

For wide sorting fields or larger material objects, the Kyptec Automation® KL-1408 8.5 MM SWIR Camera Lens can be evaluated because it provides the widest focal-length geometry in the current Kyptec Automation® SWIR family. The model is designed for 900–1700 nm imaging with F1.4 aperture, 2 MP resolution class, 2/3-inch format and C-Mount.

This type of wide FOV is useful when each object remains large enough to provide a stable material sample. If small fragments must be classified, a tighter field may provide stronger spatial purity.

Step 9: Use an Intermediate Focal Length When Wide FOV Wastes Pixels

The Kyptec Automation® KL-1410 12.5 MM SWIR Camera Lens provides a useful intermediate geometry. Its current specifications include 12.5 mm focal length, 900–1700 nm range, 2 MP, F1.4, 2/3-inch format and C-Mount.

If a broad 8.5 mm field captures large amounts of empty conveyor or irrelevant background, moving toward 12.5 mm can allocate more sensor pixels to the actual material without creating an excessively narrow field.

For automated sorting systems, this can improve segmentation and reduce the proportion of mixed product/background pixels.

Step 10: Control Mixed Pixels

Mixed pixels occur when one pixel receives signal from more than one material.

At the boundary between Material A and Material B, the measured response can be approximated as:

M(λ) = αA(λ) + (1 − α)B(λ)

where (α) represents the proportion of the pixel occupied by Material A.

If both materials contribute strongly, the pixel may not match either reference class.

The machine should therefore classify regions using sufficient pure interior pixels wherever possible. Edge pixels can be excluded, weighted differently or processed using separate boundary logic.

Mixed-pixel control becomes particularly important in recycling, textile blends, contamination inspection and fragmented-material sorting.

Step 11: Choose a Stable Background

A conveyor or support surface should not be treated as an irrelevant mechanical detail.

If the background produces SWIR responses close to one of the material classes, segmentation becomes harder. If its spectral response changes because of wear, moisture or contamination, classification can drift.

A good production background should be spectrally stable and clearly distinguishable from the target materials at the wavelengths being used.

During validation, the system should include:

clean background;

worn background;

expected contamination;

and product-edge pixels.

This helps prevent conveyor variation from being mistaken for material variation.

Step 12: Design a Reference Library Around Production Variability

A strong material library should contain multiple reference samples for each class rather than one average image.

For each material, record the expected distribution of:

spectral response;

surface appearance;

thickness;

orientation;

temperature;

moisture;

and any other variable capable of altering SWIR intensity.

The library should define not only the center of the class but also its acceptable boundaries.

This makes the system more robust because it asks whether a new observation belongs within the validated material population rather than whether it matches one perfect reference.

Step 13: Separate Calibration References From Material References

These two references serve different purposes.

Calibration references monitor the optical system itself.

Material references define the acceptable material classes.

A white or known-reflectance reference can help monitor wavelength-dependent illumination and optical throughput. A dark reference can help remove detector offset.

A material golden sample, by contrast, represents a real production material.

These should not be treated as interchangeable.

A system can remain optically calibrated while the material specification changes, or the material can remain unchanged while illumination drifts.

Step 14: Use Dark and Reference Correction Where Appropriate

A simplified corrected reflectance-style measurement can be expressed as:

R = (I_sample − I_dark) / (I_reference − I_dark)

where (I_sample) is the measured material response, (I_dark) represents detector/background offset and (I_reference) is a controlled optical reference.

The exact implementation depends on the system, but this correction demonstrates an important principle: raw pixel intensity contains both material information and instrument behaviour.

Calibration attempts to remove as much instrument variation as practical so that the remaining difference corresponds more closely to the material.

Step 15: Calibrate Across the Entire Field, Not Only the Center

Illumination intensity and optical response can vary across the FOV.

If the center of the image is brighter than the edges, the same material can produce different raw values depending on position.

That creates a dangerous failure mode in conveyor systems: material identity appears to change simply because the object moved laterally.

Reference samples should therefore be placed across several field positions.

Where needed, flat-field correction or position-dependent normalization can compensate for systematic variation.

The production threshold should be validated at the weakest field position.

Step 16: Calibration Must Include Time Drift

A material-classification system may run continuously for months.

During that period, illumination output can age, optical windows can become dirty, focus can shift and mechanical alignment can move.

Calibration should therefore not be treated as a commissioning-only activity.

A practical system can track a reference response over time and define a tolerance window. If the reference moves outside the permitted range, the machine can require maintenance or recalibration before classification reliability is compromised.

This is especially important when class separation is narrow.

Step 17: Decision Thresholds Should Be Derived From Distributions

Suppose the classification score for acceptable Material A ranges mostly from 0.75 to 0.95 while wrong Material B ranges from 0.15 to 0.40.

Placing the threshold at 0.60 creates a useful separation margin.

But if Material A ranges from 0.55 to 0.90 and Material B from 0.35 to 0.65, the classes overlap.

No threshold can create perfect separation from overlapping distributions.

This is a crucial engineering lesson: software thresholds cannot compensate indefinitely for weak optical separation.

When overlap is excessive, improve wavelength choice, illumination, FOV, sample purity or material modeling before tightening the threshold.

Step 18: Use Two Thresholds When an Uncertain Zone Is Needed

Many industrial classification systems become safer when they use an uncertainty region.

For example:

Score > 0.75 → Accept Material A

Score < 0.45 → Reject

0.45–0.75 → Uncertain / secondary inspection

This prevents borderline observations from being forced into an incorrect class.

The width of the uncertain zone should depend on the commercial cost of false acceptance and false rejection.

High-purity material sorting may justify a wider uncertainty band than a low-risk classification application.

Step 19: Set Thresholds From Business Risk, Not Just Mathematical Accuracy

The “best” threshold is not always the point that maximizes overall classification accuracy.

Suppose a wrong material entering production creates an expensive batch failure. False acceptance may therefore be far more costly than false rejection.

The threshold should then be biased toward protecting against material escape, even if more acceptable product requires secondary inspection.

A recycling sorter may instead optimize for purity and recovery simultaneously.

A production specification should explicitly state:

false-accept tolerance;

false-reject tolerance;

unknown rate;

and class-specific detection targets.

Step 20: Test New Suppliers Before Adding Them to the Approved Population

A nominally identical material from another supplier can contain different additives, pigments, processing history or recycled content.

These variations can shift the SWIR response.

A new supplier should therefore be tested against the established material library before its products are automatically treated as equivalent.

If the new material falls inside the validated approved distribution, it may be added confidently. If not, the classifier may require recalibration or a separate approved subclass.

Step 21: Kyptec Automation® KL-1412 for Controlled Classification Stations

Where the classification region is smaller and more sensor area is needed on each product, the Kyptec Automation® KL-1412 25 MM SWIR Camera Lens can provide a useful controlled-field geometry.

The 25 mm focal length can suit incoming-material verification, individual-product inspection and localized material-classification stations where extremely wide scene coverage is unnecessary.

By reducing unused surroundings, the lens can help increase the number of pixels representing the actual material region, strengthening the spatial component of the classification measurement.

Step 22: Use Longer Focal Lengths When the Material Region Is Small

For tighter inspection areas, the Kyptec Automation® KL-1414 35 MM SWIR Camera Lens provides a narrower field while remaining within the same 900–1700 nm, 2 MP, F1.4, 2/3-inch and C-Mount platform.

This can be useful when the machine needs to classify a relatively small controlled region rather than an entire wide conveyor.

For even tighter framing or greater stand-off, the Kyptec Automation® KL-1416 50 MM SWIR Camera Lens extends the portfolio to 50 mm while retaining the same core SWIR architecture.

Longer focal length improves spatial allocation of the sensor; it does not create better spectral separation. Wavelength selection must still be solved first.

Step 23: Validate at Maximum Production Speed

A material classifier developed using stationary samples may behave differently once the objects move.

Motion blur mixes neighbouring pixels and can weaken the spectral purity of small regions. Shorter exposure reduces blur but also reduces received SWIR signal.

Suppose a conveyor moves at 2 m/s and exposure is 200 µs:

2000 mm/s × 0.0002 s = 0.4 mm motion

If the minimum material region is 2 mm, 0.4 mm of motion is significant.

Final qualification should therefore use the maximum specified production speed, minimum object spacing and realistic illumination.

Step 24: Build the Acceptance Test Around Failure Cases

A production acceptance test should not consist primarily of easy reference samples.

Include:

the two most similar material classes;

minimum target size;

maximum FOV position;

worst surface condition;

maximum acceptable moisture or thickness variation;

maximum production speed;

new batches;

and deliberately unknown materials.

The system should prove that these difficult conditions remain inside the required acceptance margin.

If the hardest cases pass reliably, easy cases usually follow. The reverse is not necessarily true.

Step 25: Keep the Production Recipe Traceable

Once wavelength, exposure, aperture, focus, working distance, calibration and thresholds are approved, they form part of the validated inspection recipe.

Changing one can shift the classification distribution.

A production system should therefore document:

lens model;

focus setting;

aperture;

camera configuration;

illumination settings;

wavelength selection;

reference version;

classification algorithm version;

and threshold values.

This makes troubleshooting far easier when performance changes after maintenance or hardware replacement.

Why Kyptec Automation® Is a Strong Platform for SWIR Material Classification

The Kyptec Automation® SWIR Camera Lens collection offers a coherent set of focal lengths for the geometry stage of material-classification system design. The current family spans 8.5 mm, 12.5 mm, 25 mm, 35 mm and 50 mm, while product pages specify 900–1700 nm operation, 2 MP resolution, 2/3-inch sensor format, F1.4 aperture and C-Mount. Kyptec Automation® also specifically identifies material identification and industrial quality control among the intended applications of its SWIR optics.

This portfolio is especially useful for OEMs because material-classification machines can require very different fields of view. Wide sorting systems can prioritize coverage, incoming-material stations can prioritize controlled sampling, and tightly framed verification systems can prioritize pixels on a small target. Kyptec Automation® allows those geometries to be addressed within one dedicated SWIR Camera Lens family rather than forcing a single focal length onto every application.

Frequently Asked Questions About Designing a SWIR Material Classification System

1. What should I test first when developing a SWIR material classification system?

Begin with representative samples of the materials that must be separated and determine whether they produce repeatable differences across the usable SWIR wavelength range. Do this before choosing the final focal length or mechanical camera location. If the required materials cannot be separated spectrally under realistic variation, optical geometry alone will not solve the problem.

2. How many samples should be used to create each material reference class?

There is no universal number because industrial variability differs substantially between applications. Use enough independently sourced samples to represent supplier, batch, thickness, colour, surface condition, moisture and other relevant production changes. The objective is to characterize the class distribution rather than collect many nearly identical images of one specimen.

3. How do I know whether one wavelength is enough for material classification?

Measure all required classes at candidate wavelengths and compare their distributions. If one wavelength creates stable separation between every important class under normal production variation, a single-band solution may be sufficient. If classes overlap, additional wavelengths or ratios may provide stronger discrimination.

4. Should I choose the wavelength where the two materials look most different?

Not necessarily. The strongest visual difference between two specimens may not remain stable across all production samples. Select the wavelength that gives the best separation between complete class populations while retaining useful signal-to-noise ratio and production repeatability.

5. What is the hardest material pair in a multi-class system?

It is the pair whose spectral distributions have the smallest separation under realistic operating conditions. That pair should influence wavelength and threshold design because it is more likely to generate classification errors than classes that are already widely separated.

6. How do I calculate whether my field of view is too wide for classification?

Divide the physical inspection width by the horizontal pixel count to obtain object-side millimetres per pixel. Then divide the minimum classification-region size by this value. If the region occupies too few pixels after accounting for blur and edge mixing, reduce the FOV, increase spatial resolution or redesign the station.

7. Why do material boundaries cause classification errors?

Pixels at boundaries can contain optical information from two materials or from product and background. Their measured signature therefore lies between pure classes. Using interior regions, dedicated boundary logic or minimum-region-size rules can reduce these mixed-pixel errors.

8. How often should reference calibration be repeated?

Calibration frequency should be determined from measured system drift rather than an arbitrary schedule. Illumination aging, contamination, optical-window changes and mechanical movement can affect the measurement. Monitor reference response over time and recalibrate before drift becomes large enough to compromise classification margins.

9. What is the difference between a calibration reference and a golden material sample?

A calibration reference checks the optical system, helping identify changes in illumination, detector offset or field response. A golden material sample represents an approved production material. Both are useful, but they answer different questions and should be maintained separately.

10. How do I choose the accept/reject threshold for SWIR material classification?

Measure score distributions for approved and rejected materials using independent validation samples. Set the threshold according to the required balance between false acceptance and false rejection. When the populations overlap substantially, improve the optical or classification design rather than relying only on aggressive threshold adjustment.

11. Should a SWIR classifier always have an unknown-material category?

It is strongly advisable when unexpected materials can occur. Without an unknown state, a completely new material may be forced into whichever known class is closest. An explicit uncertain or unknown category can provide safer handling of observations outside the validated material populations.

12. What makes the Kyptec Automation® KL-1408 suitable for wide material-classification fields?

The Kyptec Automation® KL-1408 8.5 MM SWIR Camera Lens provides the widest focal-length geometry in the current portfolio. It can be evaluated for broad sorting fields or larger material objects where sufficient pixels remain available per target. Minimum object size should always be checked at the final FOV.

13. When should an OEM move from a 12.5 mm to a 25 mm SWIR lens?

Move toward a tighter 25 mm field when the existing scene contains unnecessary surroundings or when small material regions need more pixels for stable classification. The Kyptec Automation® KL-1412 25 MM SWIR Camera Lens can therefore be useful for controlled individual-material or incoming-verification stations.

14. Can changing the lens require recalibrating the material classifier?

Yes. A lens change can alter field of view, optical throughput, spatial sampling, focus and field response. Even if both lenses cover 900–1700 nm, the complete measurement chain has changed. Reference calibration and material-validation checks should therefore be repeated after a significant optical change.

15. Why does classification accuracy sometimes fall near the edge of the image?

Illumination uniformity, lens field response, perspective and reduced spatial sampling can all influence edge performance. Test equivalent materials across the complete field and apply appropriate flat-field or geometric correction where necessary. Production acceptance should use the weakest validated image region.

16. How should a new supplier material be added to an existing SWIR classification system?

First measure several independently sourced samples from the new supplier using the existing production setup. Compare their distribution with the approved class. If they fit comfortably inside the validated population, they can potentially be incorporated. If they form a shifted distribution, the material model should be updated and independently revalidated.

17. Can SWIR material classification work when the same material is available in many colours?

Potentially yes, because SWIR classification can use material-dependent response beyond visible colour. However, pigments and additives can themselves influence SWIR behaviour. Every approved colour or formulation should therefore be represented during development instead of assuming visible-colour independence automatically guarantees spectral equivalence.

18. What should be included in a production acceptance test for a SWIR material classifier?

Include the most similar material pair, smallest target, edge-of-field positions, normal supplier and batch variation, maximum production speed, realistic surface conditions and samples outside the approved library. Measure false accepts, false rejects and uncertain classifications rather than reporting only overall accuracy.

19. What information should I send before buying a SWIR camera lens for a material classification machine?

Provide material classes, sensor size, inspection width, working distance, smallest material region, product speed, likely wavelength range, number of objects in the field and whether the system needs exact classification or only approved-versus-abnormal detection. These details allow focal length and FOV to be selected from the real inspection requirement.

20. Why is Kyptec Automation® a strong choice for SWIR material classification optics?

Kyptec Automation® provides a dedicated SWIR Camera Lens collection covering 8.5 mm, 12.5 mm, 25 mm, 35 mm and 50 mm focal lengths within a consistent 900–1700 nm platform. Current models are specified around 2 MP resolution, 2/3-inch sensor format, F1.4 aperture and C-Mount. This gives machine builders a practical range for moving from broad sorting fields to tightly controlled material-verification stations without leaving the same specialized SWIR optical family.

Conclusion

Designing a successful SWIR material classification system requires the optical, statistical and production decisions to be made in the correct order. The process should begin with verified material populations and the exact production decision that must be made. Candidate wavelengths should then be compared according to how reliably they separate those populations, especially the most difficult material pair. Only after useful spectral contrast has been established should the machine builder calculate minimum target size, object-side pixel sampling, working distance and required field of view.

Reference-sample design is equally important. Industrial materials do not exist as one perfect spectrum, so supplier, batch, moisture, surface finish, thickness and other legitimate variation must be represented. Calibration should then separate instrument behaviour from material behaviour through appropriate optical references, field correction and drift monitoring. Decision thresholds should be derived from validated class distributions rather than selected manually from a few demonstration images, and an uncertain or unknown category should be included where the production risk justifies it.

The Kyptec Automation® SWIR Camera Lens collection provides a strong optical foundation once these material requirements are known. The Kyptec Automation® KL-1408 can support broad classification fields, the Kyptec Automation® KL-1410 can provide a useful balance between coverage and material sampling, the Kyptec Automation® KL-1412 can concentrate the sensor more strongly on individual products, while the Kyptec Automation® KL-1414 and Kyptec Automation® KL-1416 can support progressively tighter inspection regions and greater stand-off requirements. Across the portfolio, the shared 900–1700 nm orientation, 2 MP resolution class, 2/3-inch format, F1.4 aperture and C-Mount provide OEMs with a coherent platform for building material-classification machines around verified spectral behaviour.

For industrial buyers and system designers, the strongest rule is therefore: prove the material separation first, preserve enough pixels on the smallest target, calibrate the optical chain, and set thresholds from real production distributions rather than ideal samples. When wavelength selection, lens FOV, reference libraries, calibration and decision logic are engineered as one system, Kyptec Automation® SWIR Camera Lenses can provide a technically strong basis for automated material identification, quality verification and industrial classification across the 900–1700 nm SWIR range.