If a defect has no contrast in the wavelength band your sensor sees, it is a physics problem, not an AI problem, and no amount of training data fixes it. Teams miss this and spend months adding images to a model that is being asked to learn a distinction the photons never carried. The model cannot see what the sensor never captured.
The reframe is simple and expensive to ignore. A model learns to separate classes from the differences present in its input. If moisture, subsurface contamination, a coating-thickness change or a polymer swap produces no measurable contrast in visible RGB, then the training set contains no signal to learn from, however large it grows. The fix is not a better model. It is a band where the difference exists.
Where defects become visible by band
Different materials and defects discriminate in different parts of the spectrum. Choosing the band is the inspection decision; the model comes after.
| Band | Reveals | Typical examples |
|---|---|---|
| Visible (RGB) | Colour and gross surface features | Scratches, print defects, obvious discoloration |
| NIR | Slight subsurface and moisture differences | Bruising under skin, some coating variation |
| SWIR | Moisture, water content, many polymer and coating differences | Wet vs dry regions, subsurface stains, plastic sorting, fill level through packaging |
| UV | Surface fluorescence and some contamination | Adhesive coverage, certain residues, coating gaps |
| Thermal | Emissivity and heat differences | Seal integrity by temperature, some subsurface voids |
The pattern worth internalising: moisture, subsurface contamination and polymer differences frequently sit outside the visible band, which is exactly the set of "invisible" defects teams keep blaming on the model.
Walk one band-selection decision end to end
Suppose a line misses a subsurface moisture defect that is invisible in RGB.
- Defect-to-material contrast — the defect is water content; water absorbs strongly in SWIR, so contrast exists there and not in visible.
- Band — SWIR is the candidate.
- Sensor cost delta — a SWIR sensor is a real step up in cost from a monochrome or RGB industrial camera, so the band choice has to be justified by the defect, not by curiosity.
- What the model can finally see — with SWIR contrast present, the moisture defect becomes a separable class, and a modest model learns it. The expensive sensor earned its place by carrying the signal; the model only had to classify it.
Run that chain before buying anything. The band is chosen by the physics of the defect, not by the sensor you already own.
The two classic failures
- Spending on models for invisible defects. Months of labelling and retraining against a defect with no contrast in the captured band. The symptom reads as "the model is not good enough". The cause is a sensor that never carried the information.
- Spending on SWIR for a problem a polarizer solves. The opposite error: reaching for an expensive multispectral rig when a glare or specular-reflection problem would yield to a polarizing filter costing a fraction of it. Not every invisible defect needs a new band; some need better control of the light you have.
Symptom to cause, so you blame the right thing
| What you see | Common misread | Actual cause |
|---|---|---|
| Defect blurry or smeared | "Model needs more data" | Exposure or motion blur — a settings problem |
| Defect too small to resolve | "Model needs more data" | Pixels per defect — a resolution problem |
| Defect simply not present in image | "Model needs more data" | No contrast in the captured band — a wavelength problem |
Only the third row is a band problem. The discipline is to diagnose which of the three you have before spending, because the three fixes are unrelated and only one of them is a new sensor.
The honest boundary
This page is physics, not a Hypernology multispectral case study. The band-contrast tables above are optics reasoning; Hypernology's deployed proof is in learning-based defect detection on visible and monochrome industrial imaging at volume. The transferable point is the order of operations: fix the physics so the signal is in the image, then let the model classify it. A platform that learns defects efficiently still cannot learn one the sensor never recorded.
POV restated: the model cannot learn a distinction the photons never carry. Choose the band first.
Related reading: the lighting and illumination guide, the imaging components guide, and the defect-detection fundamentals guide.
Frequently asked questions
Why can't my model detect a defect I can barely see? If the defect has low or no contrast in your sensor's band, the image contains no signal to learn from. More training data cannot create contrast that the wavelength did not capture.
When is SWIR worth the cost? When the defect is moisture, water content, or a polymer or coating difference that produces contrast in SWIR and none in visible. If the problem is glare or specular reflection, a polarizer is far cheaper and often sufficient.
How do I know if it's a band problem or a model problem? Check whether the defect is visible to a careful human in the captured image. If a person can see it and the model cannot, it is likely a model or resolution problem. If no one can see it in that band, it is a wavelength problem.
Send examples of the defect you cannot catch, with the material and the lighting you use now, and within two weeks we will return which band carries the contrast and the sensor-versus-polarizer decision before you spend on either. Start at https://apac.hypernology.net/contact.