A 300-page book arrives as a PDF, but no word can be selected, searched, or counted. The pages look readable to a person; to translation software, they’re pictures. Starting translation before converting those pictures into checked text can turn a small recognition error into a repeated error across the manuscript.
A scanned-book translation project has two connected jobs: recover the source text from the page images, then translate and review that text without losing the book’s structure. OCR, or optical character recognition, identifies printed characters in an image and makes them machine-readable. OCR is not a substitute for proofreading: recognition errors can pass into later text formats and then into the translation.
What a scanned-book translation workflow has to solve¶
A scanned book is a set of page images, sometimes placed inside a PDF. A text-based PDF contains selectable characters; an image-only PDF does not. When a translator copies from an image-only file, there’s no actual text to copy. OCR must create a text layer before a text-based translation workflow can begin. Internet Archive’s troubleshooting guidance also warns that OCR can be poor or garbled.
According to the Internet Archive Help Center:
“OCR (Optical Character Recognition) is inexact. Sometimes it can be poor.”
That short warning has a practical consequence: a clean-looking OCR document isn’t proof that every word is right. A recognition error can alter a name, a negation, a number, a footnote reference, or a specialist term. Once the source text moves into a translation memory, a CAT tool, an EPUB, or a plain-text file, a reviewer may no longer have the page image beside it.
A book also isn’t one uniform block of prose. A novel may contain chapter headings, dialogue, quotations, page headers, and footnotes. A technical manual may add tables, equations, labels, diagrams, and captions. Historical books may include old type, faded ink, marginal notes, or multiple scripts. Each page type creates a different recognition and layout problem.
Treat OCR quality, source-text correction, translation, and layout reconstruction as separate checkpoints. A team that combines them into one vague “conversion” task can’t see where errors entered the work or assign an owner to fix them.
A useful project handoff includes the original images, the OCR output, the corrected source text, the translated text, and a way to compare each stage. When layout matters, keep page references or coordinates so reviewers can trace a sentence back to its place in the source. Internet Archive’s OCR documentation describes hOCR, a structured format that retains page text and bounding boxes.
The first decision isn’t which translation tool to use. It’s whether the source images are fit for recognition, whether the book’s language and page sequence are known, and which parts need human attention. Those answers determine whether a team should process the whole book, rescan selected pages, or pause for a manual transcription plan.
Prepare page images before running OCR¶
OCR reads pixels, not the book’s meaning. A pale letter on a stained page, a rotated line, or a dark gutter can make one character look like another. Page orientation, cleanliness, color, typeface, and language all affect output, according to Internet Archive’s OCR troubleshooting page.
Resolution needs context. Internet Archive says it scans books at 600 dpi, while its book-image upload guidance recommends at least 300 dpi. Those figures describe different contexts, not one universal minimum for every private book or OCR engine. The Internet Archive guidance is the source for both figures. For cultural heritage work, the FADGI third edition, published in May 2023, lists bound-volume performance levels from at least 190 ppi to at least 396 ppi. The project should choose capture quality for its preservation and use requirements, not copy a number without checking what it applies to.
A separate U.S. National Archives requirement applies to covered federal permanent-record digitization, not to every privately scanned book. Its guidance, published on June 7, 2023, specifies at least 300 pixels per inch for modern textual paper records and at least 400 ppi for photographic prints and paper records with fine details. The same guidance calls for RGB color when color or other characteristics matter to interpretation. Read the National Archives explanation before applying those requirements to a project.
Check the image set before recognition. Internet Archive accepts JP2, JPG/JPEG, TIF/TIFF, GIF, BMP, and PNG for book-image uploads, and asks for crisp images in sequential page order. Its upload instructions also say to include covers, check for skipped pages, and make page lines horizontal, or vertical for some languages.
| Check before OCR | What to inspect | Why the check matters |
|---|---|---|
| Page sequence | Covers, front matter, numbered pages, inserts, and the final page | A skipped scan can make the translation incomplete even when each supplied image is clear. |
| Orientation | Text lines run in the expected reading direction | Rotated pages can produce garbled recognition or a wrong reading order. |
| Image clarity | Focus, contrast, shadows, stains, and the inner gutter | Faint or obscured characters give OCR less visual information. |
| Language | Set the main language and note any additional scripts | Language selection affects which characters the recognizer expects. |
| Page types | Include prose, headings, footnotes, tables, captions, and illustrations | A sample of only ordinary paragraphs won’t reveal how other elements behave. |
| File handling | Keep an unchanged copy of the original images | Reviewers need a stable reference when correcting OCR output. |
Language metadata is not a minor setting. Internet Archive’s developer documentation says its OCR workflow uses item language metadata to determine OCR languages. Multiple language values may be provided, and missing or unknown language information can trigger autonomous detection behavior. See the OCR documentation.
A book that switches languages needs more than a single label in a project note. Record which pages, chapters, or elements use each language, then test whether the OCR tool reads them correctly. A foreign-language quotation inside an English chapter can be misread if the recognition setup assumes the page contains only English.
The Internet Archive upload guide says:
“All images are crisp.”
Crispness is a useful upload check, but it doesn’t mean the text has been recognized correctly. Image quality is an input condition; the OCR output still needs review.
Choose an OCR and translation setup that fits the job¶
Translation teams usually weigh three routes: a managed book-digitization or translation bureau, a freelance OCR or translation specialist, or an online document workflow. The right choice depends on source condition, layout demands, review capacity, and the deliverable. A clean text-only novel and a fragile illustrated volume shouldn’t automatically receive the same workflow.
| Route | Useful when | What the team still needs to control |
|---|---|---|
| Bureau or digitization provider | The project needs scanning, OCR, text correction, translation, or page reconstruction coordinated as services | Define who approves OCR corrections, how page-level queries are handled, and which files are delivered at each stage. |
| Freelance specialist | The team can split OCR, proofreading, translation, and layout work among named contributors | Make sure the handoffs preserve page references, terminology decisions, and responsibility for corrections. |
| Online document workflow | The images are usable and the team wants a working document to review and post-edit | Confirm file support, language coverage, privacy needs, and limits for degraded or handwritten material. |
A bureau can coordinate several disciplines, but a quote should say whether it includes OCR correction and layout work or only text translation. Ask for a sample deliverable and a clear definition of “corrected.” That word could mean fixing obvious character recognition mistakes, checking every line against the scan, or rebuilding page design; those are different scopes.
A freelancer can be a good fit when an agency already has a translator or editor and needs a specific task, such as OCR cleanup for a difficult chapter. The project lead still needs to make sure the person receives page images, not just a plain-text export. A text-only file hides the original placement of headings and footnotes and can make source verification slower.
Online tools can reduce manual retyping, but not every tool handles image-only files or preserves layout in the same way. A typical situation is a translator who needs an editable draft from a scan, then wants to review the translation with the original visible. In ChatsControl, the scanned-document workflow reads a scan or photo, builds an editable Word copy that follows the page layout, and then translates the Word document. A separate check compares the reconstructed copy with the page, and words still found missing are counted for the reader. Handwriting and very degraded scans remain unreliable, so those files need human review against the original.
A book may need a dedicated OCR engine, a CAT tool, a layout specialist, or a combination. A document-translation tool is not a full CAT tool or TMS: it doesn’t replace translation memory, segment-level project management, or team vendor workflows. Choose tools around the work you actually need, rather than assuming a single application will cover scanning, OCR correction, translation, editorial review, and book production.
Google Books is not a general shortcut for a full-book translation. Google says copying or translating selected text is available only where the feature is offered, and a publisher may not have made it available for a given book. Its help page says a book may not be fully viewable because the rights owner hasn’t agreed to show the full book or because the book wasn’t fully scanned. Read Google Books’ help page before treating selected-text features as a project workflow.
Rights need a separate check from technical access. Internet Archive tells potential uploaders to proceed only if they believe they have the appropriate rights. A business project should confirm the permissions needed for translation, distribution, and publication in the relevant territories using the Internet Archive upload guidance. Access to a digital copy doesn’t by itself establish permission to translate or publish a whole book.
Run a sample-first workflow, then scale up¶
A representative sample catches project problems while the team can still change the plan. Don’t approve an OCR method based only on a clean opening page: inspect a mixture of ordinary text, page furniture, complex layout, and difficult printing before committing the full manuscript.
- Inventory the source. Record the image or PDF files, source language, page sequence, available page types, and known gaps. Keep the originals unchanged.
- Check image fitness. Review focus, contrast, orientation, page edges, and the gutter. Mark pages that need rescanning or manual handling.
- Set the OCR language. Specify the main language and note other scripts or languages found in the book. Confirm how the chosen workflow handles mixed-language pages.
- Select a representative sample. Include body text, headings, footnotes, tables, illustrations, unusual scripts, and any degraded or rotated pages.
- Run OCR and compare it with the images. Review recognized words against the source page, not just against the output file. Record recurring mistakes and the pages or elements where they occur.
- Decide the correction threshold. Agree who can fix obvious OCR mistakes, who approves an uncertain reading, and when the team must ask the rights holder or editor rather than silently changing the source.
- Process the rest of the book. Apply the approved settings, maintain page order, and record exceptions rather than hiding them in the text.
- Translate and review. Translate corrected source text, check the translation against the original, and verify that references, tables, captions, and page structure remain usable.
For a business-scale project, this sample-first sequence is a workflow recommendation based on documented OCR variability, not a universal formal requirement. Stanford Libraries describes the OCR in its Google-scanned book collection as uncorrected, which is a reminder that available text may be a starting point rather than an edited transcript. See the Stanford Libraries guide.
Set an acceptance rule before the sample review. For instance, decide what happens when a page contains a doubtful name, a missing line, or an unreadable footnote. The team can log the page, preserve the image, and request a human reading instead of guessing. A translation brief can also record preferred spellings and terminology so names and recurring terms don’t drift from chapter to chapter.
When the project uses an editable document workflow, compare the source and translation side by side during post-editing. ChatsControl provides a bilingual review view in which the source and translation appear together, with translated text highlighted and the original available on hover. That view helps a reviewer inspect meaning and wording, but it doesn’t remove the need to check uncertain OCR against the scan itself.
What should a reviewer inspect first?¶
Start with characters whose errors can change meaning: personal names, place names, numbers, dates, negations, abbreviations, and technical terms. Then check line and paragraph boundaries. An OCR system can join adjacent columns, separate a hyphenated word incorrectly, or place a footnote inside the main paragraph.
Tables deserve a separate pass. Confirm that labels remain attached to the right values and that rows still read in the intended order. In a two-column page, verify that the OCR hasn’t read across columns. For a figure, check that captions and labels have not been detached from the image they describe.
Book layout and book structure aren’t identical. A translation may not need to recreate every line break, but the team still needs to preserve chapter hierarchy, headings, footnote links, table relationships, and page references where the publication or editorial workflow depends on them.
Keep layout and searchable text verifiable¶
OCR can produce more than a plain text file. Internet Archive’s developer documentation describes hOCR HTML with page text, bounding boxes, and word-level confidence. Character-level output may also be available in a corresponding chOCR file. The same documentation says derivative PDFs can contain an OCR result as a hidden, selectable, searchable text layer. See Internet Archive’s OCR documentation.
Bounding boxes connect recognized words to positions on the page. That matters when a reviewer asks whether a word came from the body text, a caption, or a margin note. Plain text doesn’t inherently show where the words appeared, so a team that only keeps TXT may have to reopen page images and reconstruct the context manually.
| Deliverable | What it helps with | What it doesn’t prove |
|---|---|---|
| Original page images | Visual reference for source verification | Images alone don’t give translators selectable text. |
| Plain-text OCR | Search, word processing, and text-based translation | Plain text doesn’t show where each word appeared on the page. |
| Searchable PDF | Keeps page images available with a hidden selectable text layer | Searchable doesn’t mean error-free. |
| hOCR or comparable structured output | Retains text with page coordinates and other structural information | Coordinates don’t guarantee correct reading order or correct recognition. |
| Editable Word file | Supports editing, translation, and review in a familiar document format | Rebuilt layout still needs comparison with the source page. |
Pick formats around both the translation and the final publication. FADGI’s third-edition guidance, published in May 2023, lists TIFF, JPEG 2000, and PDF/A among master-file formats for bound volumes, with capture choices dependent on project and performance level. That guidance is for cultural heritage digitization, not a rule that every translation team must use one of those formats. The FADGI guidelines help frame preservation choices separately from the working files needed for translation.
Think about the reader of each deliverable. An editor may need a clean Word file; a translator may need OCR text plus page images; an archive may require a preservation master; a production team may need a reconstructed layout. A single file can rarely answer every need without trade-offs.
For text-heavy books, a searchable PDF paired with corrected editable text can make review practical. For books with complex design, keep a structured OCR result and page images through layout work. For a simple working draft, a translated Word file may be enough, provided the team keeps a route back to the original page when questions arise.
Common failure points and how to contain them¶
Most avoidable errors come from treating OCR as a one-click conversion. The software recognizes visual patterns; it doesn’t know that a character sequence is a person’s name, that a footnote marker belongs to a sentence, or that a faded word should remain uncertain rather than be guessed.
Wrong language settings. The recognizer may misread characters if the language metadata doesn’t match the page. Internet Archive’s OCR documentation explains that item language metadata determines OCR languages, and its upload guidance says language selection affects OCR. Set the language intentionally and test any mixed-language content before the full run.
Skipped pages or covers. A clean OCR file can still be incomplete. Compare the scanned image sequence with the book’s page sequence, including covers and front matter. Internet Archive’s upload instructions explicitly say to include all pages and check for skipped ones.
Rotated or poorly framed images. A small orientation error changes the recognition task; a gutter shadow can hide letters. Check whether the lines are horizontal or, for some languages, vertical, and mark pages where curvature or cropping obscures text. The Internet Archive upload guide calls for crisp images and correctly oriented page lines.
Plain text mistaken for verified text. A searchable file only means a text layer exists. Stanford Libraries explicitly describes its Google-scanned book OCR as uncorrected. Keep the scan available until the source text has been reviewed, as explained in the Stanford guide.
Layout flattened into an unreadable stream. Plain text can lose the relationship between columns, notes, captions, and tables. Check structured output or compare the editable draft against the image. If the layout carries meaning, assign a person to verify that each element remains attached to the right text.
Silent edits to the source. A translator may find an apparent typo and be tempted to correct it. Separate OCR correction from editorial correction: record what the scan shows, note proposed changes, and get approval for changes that alter the source. A translation should not quietly turn an uncertain reading into a definite claim.
Rights assumed from availability. A book’s presence in a platform or a selected-text feature doesn’t establish permission to translate and publish the whole work. Google Books explains that viewability and copying features depend on availability and publisher choices; Internet Archive tells uploaders to proceed only if they believe they have appropriate rights. Check the Google Books help page and Internet Archive upload guidance, then confirm project permissions for the intended territories.
Stanford Libraries describes its collection’s OCR as:
“The data consists of uncorrected OCR text”
Uncorrected text can still be useful for searching and drafting, but it should not be treated as a reliable final source without review. Agree on a correction and approval path before translators inherit the file.
Plan handoffs, permissions, and regional requirements¶
A book translation can involve different teams for scanning, OCR, proofreading, translation, editing, and page design. The location of those teams doesn’t change the core technical checks: preserve the original, define the source language, check sample pages, and keep the output traceable to the image. What changes by project is the required preservation format, publication workflow, and rights clearance.
For a preservation-focused collection, capture and master-file choices may follow a cultural heritage standard. FADGI’s third edition, published in May 2023, gives performance levels for bound volumes and lists TIFF, JPEG 2000, and PDF/A among master-file formats. Its recommendations depend on project and performance level; see the FADGI guidelines.
For covered U.S. federal permanent-record digitization, the National Archives guidance describes requirements under 36 CFR 1236.50. The published guidance specifies image resolution and color requirements for the covered records; it shouldn’t be presented as a general rule for private publishing. Check the National Archives post when a project falls within that scope.
For online research collections, access rules can also shape who can work with the files. Stanford’s guide says that team members accessing its specific research collection must sign a Data Use Agreement and that access is limited to eligible Stanford affiliates, with undergraduates able to request faculty-sponsored access. Those terms apply to that collection, not to all scanned books; consult the Stanford Libraries page before planning access around it.
A project handoff sheet can prevent the most common misunderstandings:
- Identify the party responsible for image quality and page order.
- Name the person who approves OCR corrections and uncertain readings.
- State whether the translator receives page images, plain text, structured OCR, or all three.
- Record the required delivery format for translation review and final publication.
- Agree on how to flag missing, unreadable, or ambiguous passages.
- Confirm permissions for translation, distribution, and publication in the relevant territories.
Don’t promise a turnaround based only on a platform’s processing estimate. Internet Archive says that processing uploaded book images into derived formats can take from minutes to hours depending on conditions. That estimate is specific to its platform and isn’t a general OCR or translation deadline; see its upload guidance. A book project’s schedule also includes sample review, correction, translation, editorial decisions, and layout.
FAQ¶
How do you translate a scanned book that has no selectable text?¶
Run OCR first so the page images become machine-readable text. Compare the OCR output with the scans, correct or flag uncertain readings, translate the checked source, and review the result against the original pages.
What scan resolution is best for OCR on book pages?¶
No single threshold applies to every book and project. Internet Archive recommends at least 300 dpi for book-image uploads and says it scans books at 600 dpi for its own operation; those figures refer to different contexts in its OCR troubleshooting guidance.
How should a translation team quality-check OCR before translating a whole book?¶
Process a representative sample that covers body text, headings, footnotes, tables, illustrations, unusual scripts, and poor-quality pages. Compare recognized text with the images, document recurring errors, and agree on who reviews uncertain readings before running the full book.
Can OCR preserve footnotes, tables, page order, and book layout?¶
Structured OCR such as hOCR can retain page coordinates, while plain text doesn’t inherently show where words appeared. Neither format by itself guarantees correct reading order or layout, so check page sequence, footnote links, table relationships, captions, and columns against the scans.
What file formats and language settings should be prepared for book OCR?¶
Internet Archive’s book-image upload guidance accepts JP2, JPG/JPEG, TIF/TIFF, GIF, BMP, and PNG. Keep images crisp and sequential, orient the page lines correctly, and select the book language because language metadata affects OCR.
Can I use Google Books to translate an entire scanned book?¶
Google Books allows copying or translating selected text only where that feature is available, and its help page says some books may not be fully viewable. The feature doesn’t establish general permission or a general workflow for translating and publishing a whole book; check Google Books’ guidance.
Who is responsible if OCR errors reach the published translation?¶
The project team should assign responsibility for OCR review, source-text approval, translation review, and final production checks. OCR is inexact, so the handoff should preserve page images and a clear way to report uncertain or missing text rather than passing unverified output forward.