Scanned PDF to DITA XML: OCR, Repair, Convert

Yes, a scanned PDF can be converted to DITA, but not in one step. The page images go through OCR first, the document is rebuilt into real headings, lists, tables and figures, the damage the scan left behind is repaired by rule, and only then is the result converted to DITA, validated and reviewed.

Why a scan cannot be converted directly

Conversion rules read structure out of a source and map it to DITA elements. On a scanned page there is no structure to read. There is no heading element to detect, because there is no element at all: there is a picture of a heading. Until OCR runs there is no text either, and OCR output is never clean. The guide to PDF to DITA conversion covers PDFs of every kind; this page is about the scanned ones.

What scan damage looks like

Scanned books fail in a specific, repeatable way. Words are broken by imperfect character recognition or cut in half at line ends. Diagram pages arrive shredded into image fragments, with their callout labels adrift as loose text. Captions drift away from the figures they belong to. Running headers and page numbers land in the text, sometimes glued onto a heading. Tables lose their grid.

The read-only assessment comes first

Before anything is changed, the document is read and assessed: how many pages, how many images, and how this particular file is damaged. The assessment is always on and strictly read-only. It finds the broken words and the scattered figure labels, reports them, and changes nothing. Across an archive, that report is a defect map with counts: you see the condition of your pages while you still have the option to walk away.

Rebuilding the document

OCR produces text. It does not produce a document. The next step rebuilds one: inferred headings, lists that renumber when you edit them, tables reassembled with actual rows and cells, and images placed where they belong on the page. This step sets the ceiling for everything downstream, so we use a leading commercial reconstruction engine. It processes documents in its cloud, and we confirm that this is acceptable for your content before anything is uploaded.

What is repaired by rule

The everyday damage is fixed with no AI at all, by rules that give the same output for the same input:

Figures. A shredded diagram region is re-rendered whole from the source PDF. Fragments are never re-stitched. The loose callout labels are kept as the image's alt text.
Running headers and page numbers are removed, even where the scan glued them onto a heading.
Hyphenation at line breaks is healed against the document's own vocabulary, so genuine compound words survive.
Captions are normalised, and side-by-side figure strips are split so that each figure keeps its own caption.
Dot-leader listings are rebuilt into real tables.
Lost chapter headings are restored by cross-checking the book's own table of contents.

What is assisted, and how tightly

There are three AI assists, and each one is optional. The first two act only on fragments the assessment has already flagged as damaged; the third is a deeper pass that runs only on request:

Word repair. Where recognition has mangled a word, the pass transcribes that flagged word image. It sees the word with the sentence it sits in, never the document as a whole.
Figure titles. A bare figure label that carries no description can be given a short descriptive title.
Whole-text proofread. A deeper pass for badly scanned books, off unless you ask for it and restricted to exact corrections of recognition errors. It does not rephrase.

Every decision is remembered, so the same damage is never examined, or charged for, twice. The reasoning behind this split between rules and AI is set out in AI + deterministic conversion.

From repaired document to DITA

What comes out of the repair is an ordinary structured document. It goes through the same conversion, validation and delivery as content from an authoring tool, and the DITA completeness check validates the whole set against your DTD. Two rules hold throughout. Repair passes relocate or correct text; they never delete it. And because repair runs on the reconstructed document, refining a result later does not mean rebuilding the document from the PDF again.

Why scanned material is scoped separately

Content with machine-readable structure and content without it have different conversion economics, different risk, different review effort and different schedules. Put them in one work package and the predictable work inherits the timetable of the uncertain work. So we scope the scanned pile on its own: the content from authoring tools starts converting while the read-only assessment runs across the scans in parallel. With that data, each scanned document is triaged into one of four routes: re-scan, OCR plus rule-based repair, optional AI repair, or retire.

Before any of that, hunt for originals. If the Word file behind a scanned manual still exists anywhere, Word to DITA conversion is the better route. Planning a mixed archive is covered in legacy content migration, and why DITA migrations stall explains why this pile decides so many projects.

The 1925 test

To check the method on the worst input we could find, we used a service manual printed in 1925: scanned as page images, with no usable text layer and no styling information. It came out as well-formed DITA topics under a real bookmap with per-chapter submaps, with no placeholder or dummy topics standing in for content that failed to convert. OCR damage was repaired. Shredded figures were reassembled by re-rendering them from the source. AI touched only the fragments the quality pass flagged, and every decision was cached, so the passes that followed made no AI calls at all.

What review to expect

Expect scanned material to need more review than content from an authoring tool. Inferred structure needs human eyes, and review capacity, not conversion throughput, is usually the real constraint. A reviewer checks every deliverable, and every correction becomes a permanent rule, so the same issue does not come back on the next run.

Frequently asked questions

Can a scanned PDF be converted to DITA XML?

Yes, but a scan has to go through OCR first, because reconstruction rebuilds structure and cannot read pixels as text. After OCR the document is rebuilt into headings, lists, tables and figures, the scan damage is repaired, and the result is converted to DITA and validated.

Does AI rewrite the text of a scanned manual?

No. Figure reassembly and the text and structure polish are fully deterministic. Two narrowly gated AI assists run only where the read-only assessment flags damage, and both are optional: word repair, which transcribes the individual word images that recognition mangled, and short descriptive titles for bare figure labels. Neither rewrites your prose. A deeper whole-text proofread exists for badly scanned books. It is off unless you ask for it and makes only exact corrections of recognition errors; it does not rephrase.

How much does scanned PDF to DITA conversion cost?

It depends on things that can be measured: how many pages there are, how badly character recognition has damaged the text, whether the editable original still exists, and how much tailoring to your specialization and metadata model you want. That is why the read-only assessment runs first and changes nothing. The guide to DITA conversion cost covers the drivers.

Related guides

Send us 25 pages. The messier, the better.

We'll convert them to DITA free of charge — through the real pipeline, not a demo — and review the output with you. Then we'll discuss pricing one-to-one.

Get your free 25-page sample

Scanned PDF to DITA conversion