Appearance
When multispectral capture goes wrong, work through it in physical order — lighting, geometry, exposure, calibration, then processing — because a fault early in that chain masquerades as something else later. The fastest single check is to flat-field with fresh white and dark references and confirm the bands are registered; that resolves the majority of "the data looks broken" reports. This guide is the end-to-end triage I run in the imaging room, with the concrete symptom-to-fix mapping for each stage.
Why won't my bands align?
Misregistration is the most common and most damaging fault: each waveband is a separate exposure, so any movement of the camera, the page or the focus ring shifts a band by a pixel or more, and false-colour or PCA then paints colour ghosts along every stroke.
Fixes, in order of impact:
- Lock focus and aperture before the first band; never refocus mid-sequence.
- Use a rigid copy stand and a book cradle that does not creep.
- Trigger remotely to avoid touching the rig.
- Register in software afterward with
Linear Stack Alignment with SIFT(Fiji) or phase cross-correlation to absorb residual drift.
Why is one band darker or brighter than the others?
Because exposure must be set per waveband. A halogen lamp and a silicon sensor are far more sensitive in the visible than in the UV or deep IR, so a single fixed exposure clips some bands and buries others in noise. Set each band's exposure so the white reference reads near 90% of full scale without clipping:
text
per band: expose so white-ref ≈ 0.9 * max, no pixels at saturation
then: corrected = (raw - dark) / (white - dark)The flat-field divide is what makes bands comparable afterwards.
How do I get rid of glare and hotspots?
Specular reflections off parchment, gilding or a glass platen show up as blown-out patches that destroy any text underneath. The standard answer is cross-polarisation: a polarising sheet on each light, a polarising filter on the lens, rotated 90 degrees to the lights. Combined with diffuse, even illumination at ~45 degrees, this removes the mirror reflection while keeping the diffuse signal that carries information.
| Symptom | Cause | Fix |
|---|---|---|
| Bright blown patch | Specular glare | Cross-polarisation |
| Bright edge near gutter | Raking angle too steep | Lower light angle, diffuse |
| Sheen over gilding | Metallic reflection | Polarise + reposition lights |
What causes banding or a grid pattern?
Two unrelated culprits. Light flicker — cheap LED or fluorescent sources pulsing at mains frequency — beats against a short exposure to produce horizontal banding; cure it with a flicker-free source and an exposure longer than one mains cycle. Fixed-pattern sensor noise produces a static grid; subtract a dark frame captured at the same exposure and temperature to remove it.
Why do good-looking images fail in analysis?
This is the quiet killer. The image looks fine because your screen applied an automatic stretch, but the saved master was 8-bit or JPEG-compressed, throwing away the faint differences the whole technique depends on. Rules that prevent it:
text
- Save masters as 16-bit linear TIFF (no gamma, no JPEG).
- Flat-field every band against white and dark references.
- Apply contrast stretches only to derivative copies.
- Keep the raw and the calibration frames.How do I separate a real feature from an artefact?
Reproduce it. Re-capture the region and check whether the faint mark survives flat-fielding and registration and appears in the bands physics predicts. Artefacts are fragile — change the light angle, re-flat-field, or re-register and they shift or vanish. A genuine palimpsest stroke or faded line is stubbornly repeatable across captures and behaves consistently across wavebands. Document each test so the conclusion is defensible.
Key Takeaways
- Troubleshoot in physical order: lighting, geometry, exposure, calibration, processing.
- Misregistration is the top fault — lock focus, use a rigid stand, and align bands in software.
- Set exposure per band so the white reference sits near 90% without clipping.
- Kill glare with cross-polarisation and diffuse 45-degree lighting.
- Banding comes from light flicker or fixed-pattern noise; use flicker-free sources and dark-frame subtraction.
- Save 16-bit linear masters and flat-field; never analyse a JPEG or 8-bit copy.
- Validate faint features by reproducing them across re-captures and expected bands.
Frequently Asked Questions
Why are my multispectral bands not aligning?
The page or camera moved between exposures, or focus shifted. Lock focus and aperture for the whole sequence, use a sturdy copy stand and a book cradle, and register the bands in software afterwards to absorb any residual sub-pixel drift.
Why is one band much darker or brighter than the rest?
Exposure was not balanced per waveband. Each band needs its own exposure because lamps and sensor sensitivity vary with wavelength; set exposure so the white reference sits near 90% without clipping in every band, then flat-field.
What causes glare or hotspots on the page?
Specular reflection from raking or frontal light, common on glossy parchment and gilding. Use cross-polarisation (polarising filters on both lights and lens) and even, diffuse illumination at roughly 45 degrees to kill the hotspots.
Why do I see banding or a grid pattern in the image?
Usually flicker from LED or fluorescent lighting beating against the exposure, or fixed-pattern sensor noise. Use a flicker-free light source, lengthen exposure past one mains cycle, and subtract a dark frame to remove fixed-pattern noise.
My results look fine on screen but fail in analysis — why?
You probably skipped flat-fielding or saved a lossy/8-bit version. Keep 16-bit linear masters, flat-field with white and dark references, and only stretch contrast in a derivative copy, never in the master.
How do I tell a real faint feature from an artefact?
Reproduce it. Re-capture the region, check it survives flat-fielding and registration, and confirm it appears consistently across the expected bands. A true feature is repeatable; an artefact usually moves or vanishes when you change one capture variable.