A bound medieval manuscript is, from an engineering point of view, an awkward photogrammetric subject. It is fragile, it cannot be moved past about thirty degrees off its resting position without stressing the binding, and many of the ones we most want to document are charred, water-damaged or already partially detached from their boards. The standard photogrammetric workflow — put the object on a turntable, spin it slowly, fire the shutter every ten degrees — is exactly the opposite of what the conservation team will let you do.

This is the workflow that the imaging studio at the Bodleian Library and our Sydney lab have settled on after eighteen months of iteration. It has now been used on twenty-three bound manuscripts ranging from a tenth-century Boethius to a charred fragment of the Codex Tinctus Aleppensis that I am not going to name precisely for the same reason I won't name a Maya dig site by its full coordinates.

The problem in one sentence

We need to produce a sub-millimetre photogrammetric mesh of a bound page that is allowed to move by no more than a few degrees, in lighting that the conservator considers safe, in a session short enough that the manuscript doesn't acclimatise to the studio environment any more than necessary.

That last constraint is the one most digital-imaging guides skip. A medieval manuscript that has lived in a cool stack for centuries does not enjoy a four-hour photography session under hot LEDs. The shorter the session, the better.

The rig

The camera is a Phase One IQ4 150MP back on a P+ body, but the workflow generalises to any 60+ megapixel sensor with a good macro lens; we have replicated it with a Sony A7R V and a 90mm Sony macro for projects that don't justify the medium-format. The rig is built around three principles.

1. The manuscript does not move. The object is on a heavy, cradle-supported book holder built from a conservation pattern used by the Bodleian for decades. The cradle holds the page open at a maximum 110° opening angle; the page is held against a glass platen with a calibrated low-pressure foam-edged sandbag system. No mechanical contact with the page surface itself.
2. The camera moves on a programmable arm. We use a Cognisys StackShot 3X rail mounted on a Manfrotto column, programmed to traverse a 30 cm × 22 cm path in 96 positions per page, taking a five-image focus stack at each position. Every shot is taken at the same focal distance to the page surface, plus or minus 1.8 mm.
3. Lighting is cool, polarised and brief. Two cross-polarised Profoto B10 heads at low output, triggered for less than 1/100 s per exposure. The cumulative light dose on a single page during a full imaging session is below the IIC guideline of 50,000 lux-hours for an annual exposure.

Total time per page, from the conservator setting the page to the shutter going quiet: 22 minutes. That is more than the conservator would prefer and less than I would prefer, and that compromise has held up across six different institutions now.

The capture

For each page we collect 480 raw frames (96 positions × 5 focus stacks). Those frames are stacked in Helicon Focus to produce a single ultra-deep-focus composite per position, then the 96 composites are run through Agisoft Metashape as a photogrammetric set.

The Metashape parameters that matter:

Align Photos:    Accuracy = High, Generic preselection, Reference preselection (source = camera).
Optimize:        f, cx, cy, k1, k2, k3, p1, p2 — every cycle.
Dense Cloud:     Quality = High (Ultra is overkill on a 2D-ish object, and quadruples the runtime).
Mesh:            Surface = Arbitrary, Source = Depth Maps, Face count = 2 million.
Texture:         Mapping = Generic, Blending = Mosaic, Texture size = 8192 × 8192, single texture.

The "Arbitrary" surface type matters. The default "Height field" works for flat objects, but a charred page is not flat: it cockles, lifts at the edges, and sometimes has lacunae where the parchment has burned through. Arbitrary handles that.

Total reconstruction time on a single page is roughly 45 minutes on our workstation (Threadripper 5965WX, A6000, 256GB RAM). The mesh density on a typical good page is around 4.2 million faces, with a mean point-to-point distance of 0.34 mm. The texture resolution comes out to about 38 pixels per millimetre on the writing surface, which is enough to resolve individual hair-side and flesh-side pores on the parchment.

Why bother with the mesh at all?

The obvious objection — "you already have 480 raw photographs, why do you need a 3D model?" — has been raised on every project, and it has a good answer.

A 3D mesh is a far better record of the physical state of the page than a 2D photograph. It captures the cockling, the precise opening angle of the binding, the depression where the ductus of a quill has pressed into the parchment, the lift of a wax seal off the surface. For a conservator considering an intervention five years from now, the mesh is the document that lets them see exactly what the page looked like the day you photographed it — not a flattened photographic approximation.

It also lets us do something that a flat photograph cannot: differential analysis. When the same page is re-imaged in three years' time and registered against this mesh, any deformation — a new fold, a fracture along an existing crease, a loss of material at an edge — surfaces immediately as a positive or negative offset in the difference map. That is the basis on which long-term condition monitoring of a charred manuscript actually works.

What this isn't good for

It is not good for reading the text. The traditional multi-spectral imaging workflow is dramatically better at recovering faded or charred ink than any photogrammetric pipeline. Where we have both — and on the high-priority charred fragments we do — we run them as complementary modalities, with the multi-spectral capture happening first (it is faster and uses less light) and the photogrammetric capture afterwards.

It is also not a replacement for the conservator's eye. There has been no project in which the mesh found a defect that the conservator hadn't already flagged. The mesh records it; the conservator detects it. That order matters.

Where this is going next

The next experiment is whether we can recover sub-surface detail from the mesh by combining oblique cross-polarised photography with raking-light captures at a 12 degree elevation angle. Preliminary results on a thirteenth-century Sarum missal suggest we can resolve dry-point ruling lines that are invisible in both the standard photographic record and the existing reflectance transformation imaging (RTI) capture. That would, if it holds up, give us a workflow that is non-contact, sub-millimetre, and reveals a class of evidence that has previously required actual surface raking with a fibre-optic probe.

The current parameter sweep is running. If it works I will publish the side-by-side comparisons in the autumn.

If you are a manuscripts conservator and you want a copy of the rig blueprints and the Metashape project template, write to [email protected] with your institutional affiliation and I will send them over.

— Elara