How to Convert Scanned Book to Text Without Losing Pages

To convert a scanned book to text, scan one page per image in page order, run OCR on the set, and export Word or plain text with each passage tied to its printed page number. Bound pages read worse near the spine, so check line ends there, and count pages against the volume before you rely on the result.

Can a scanned book be converted to text you can search and quote?

Yes. A book printed in ordinary type converts into text you can search, copy and quote, and with some care each passage stays tied to the page it came from. Expect clean results in the middle of each page and weaker ones at the inner margin, in footnotes and wherever the page was not flat on the glass.

The work has four stages, whether you want to scan a book and convert it to Word or only pull a few pages of text out of it:

1. Capture. Photograph or scan the pages, one page per image, in order, at a resolution that keeps the smallest type sharp.
2. Prepare. Split any two-page images, straighten each page, and put the files in reading order with names that sort correctly.
3. Convert. Run recognition and export in the format that suits the use.
4. Verify. Count the pages, then read what you will quote against the print.

Pick the export by what you plan to do with the text.

Size limits decide how you divide the volume

A scanned book is large, and tools set limits. Schema Law accepts a PDF, PNG or JPEG of up to 20 MB, and a color scan of a few hundred pages will exceed that as one file. Page limits apply too: three pages per document on the free plan, 250 on Solo and 1,000 on Team, as the pricing page lists. Divide the volume into parts at chapter or signature boundaries, and name each part with the printed pages it covers, for example minute-book-3_pp-101-150.pdf. A name like that makes a missing range obvious.

Test before you scan the whole thing. Three pages from the middle of the book, uploaded on the free plan, show how the type, the paper and your capture method come through. Recognition is strongest on clean print, and nothing marks a passage as uncertain, so compare the parsed pages with the scans yourself.

Why do bound pages read worse than loose sheets?

A loose sheet lies flat and lit evenly. A bound page is held at one edge, so it rises toward the spine, picks up a shadow there, and brings its text close to the fold. Recognition suffers along that inner strip of every page while the outer two thirds read normally.

Gutter shadow

Where the page lifts off the glass, less light returns to the scanner. The inner margin darkens from white to gray to near black. Letters that fall inside the gray lose contrast, and a black-and-white scan may turn the whole strip into a solid band that swallows the first or last word of each line. Scan a bound volume in grayscale or color so the shadow stays a shade of gray with letters visible in it.

Curvature

Lines of type bend as they approach the spine, and the letters there are narrower because they are seen at an angle. No rotation fixes a line that curves. The post on crooked scans explains what warp does to recognition and what dewarping can and cannot repair, under curved pages from bound volumes.

Tight inner margins

Some volumes are printed or rebound with very little inner margin. A rebound record book may have lost part of that margin to the new binding, and a thick book with a stiff spine does not open fully. The text is physically in the fold. A capture method that needs the book pressed open to 180 degrees will not reach it.

Other faults a bound book adds

• Show-through. Thin paper lets the reverse side print through as gray, mirrored text. A sheet of black paper slipped behind the page you are scanning evens out the background.
• Fingers and weights. Whatever holds the page down appears in the image and is read as shapes or characters. Hold pages at the outer margin, away from text.
• Page edges of the block. The stacked edges of the other pages show as fine parallel lines beside the page, which can be read as a column of "l" or "1".
• Uneven focus on a camera. A photographed page that curves is sharp where the lens focused and soft elsewhere.

Choosing how to capture

Do not force a binding flat to get a better scan. A record book that belongs to a client, a court or a county office is not yours to damage, and a cracked spine cannot be undone. Pages wider than your scanner are a different problem, covered in the guide to scanning legal size documents under ledger sheets, long receipts and other pages that do not fit the scanner glass.

Should you split a two-page spread before running OCR on a book?

Yes. Give recognition one printed page per image. A spread read as a single wide page invites the software to run each line straight across the fold, joining the left page to the right, and it leaves you with one image standing for two page numbers.

What goes wrong when a spread stays whole

• Line one of the left page is followed by line one of the right page, all the way down.
• The gutter shadow in the middle is read as a rule or a column of stray characters.
• The two pages lean in opposite directions, so no single rotation straightens both.
• Image 37 of the file holds printed pages 72 and 73, and every count and citation that depends on "one image, one page" is off.

The reading-order failure is the same one that affects multi-column pages, which the post on whether OCR can handle two-column documents explains.

Splitting without scrambling the order

1. Cut each spread at the fold, a little to each side. Keep the gutter in both halves so no line end is trimmed.
2. Name the halves so they sort in reading order. Zero-padded numbers work: 0072.png, 0073.png. If your splitting tool names files 37_L and 37_R, check that your system sorts L before R. For a book read right to left, the right half comes first.
3. Straighten each half on its own.
4. Use the printed page number in the file name where the book has one, and a sequence number with a note where it does not.
5. Check the pairs. In the usual Western arrangement a right-hand page carries an odd number and the left-hand page an even one. A run of files where that flips tells you a page was skipped or two halves were swapped. Confirm the pattern on your own volume first, since inserted plates and renumbered record books break it.
6. Assemble the parts. Combine the single pages into PDFs of a size your tool accepts.

Pages that are not part of the pattern

Front matter numbered in roman numerals, unnumbered plates, fold-out maps, blank versos and tipped-in sheets all sit outside the main page sequence. Scan them where they fall, and give each a name that says what it is, such as 0112a_plate-facing-112.png. Scan blank pages too. A blank page that is present in your file proves it was blank; a blank page you skipped looks later like a missing page.

Running heads, printed page numbers and footnotes in a converted book

Running heads, page numbers and footnotes are part of each page, and recognition reads them as text along with the body. In an export that flows the pages together they land in the middle of sentences that cross a page break. Keep them when you need to cite by page, and clear them only from a copy meant for reading.

What each one does

PDF page and printed page are different numbers

The fourteenth image in your file may be printed page 2, because the title page, contents and preface came first. Every citation uses the printed number. Decide how the working copy will carry it. A reliable form is a marker on its own line at the start of each page:

[p. 214]

Take the number from the page image, not from the converted text, when the two disagree. Digits in small type at a page corner are among the characters recognition misreads, and a wrong page number spoils a citation without anyone noticing. Formats that keep pages apart help here. Schema Law's HTML export is one web page with a section for each source page, and its CSV export gives the text line by line with page numbers; both let you see where one scanned page ends and the next begins. Every plan includes both, along with the other six export formats. The page numbers in an export are positions in the file, so map them to printed numbers yourself. The list of export formats describes each one.

Footnote digits inside sentences

A superscript 3 after the word "trustee" can come out as "trustee3", "trustee 3" or simply "trustee". In a statute or contract quoted in the book, a stray digit attached to a number is worse: "section 12" followed by footnote mark 4 becomes "section 124". When you copy a passage, look at the image for every digit that sits at the end of a word or number.

Quoting across a page break

A sentence that starts on page 214 and ends on page 215 has, between its halves in the raw text, the footnotes of page 214, the page number, and the running head of page 215. For a quotation:

1. Find both halves on the page images.
2. Join them in your working copy, removing the head, number and notes from between them.
3. Rejoin a word hyphenated across the break, after checking whether the hyphen belongs to the word.
4. Cite both pages in the form your citation rules require.

A reading copy and a citing copy

Keep two exports. The citing copy is the searchable PDF or a page-marked text file with heads and numbers intact. The reading copy is a Word file where you have removed repeated heads and moved notes out of the way. Never cite from the reading copy alone.

Deed books, minute books and treatises a law office scans

The bound volumes a law office scans fall into a few types, and each has its own weak point: handwriting and marginal entries in older record books, signatures and pasted-in pages in minute books, and footnotes and dense citations in treatises and reporters.

Record books

Before you scan a volume that belongs to a public office, ask the office what it permits on handling, copying and who may operate equipment. Capture the spine label, the book number and any index pages along with the entries you came for. Handwritten entries are a separate subject; the post on OCR for handwritten notes sets out what recognition can do with them.

In a deed book, an entry may grant a life estate and leave a tenancy in common in the rest, or it may mention an appurtenance such as a spring or a lane. You read that estate in land from the image at the inner margin, where the binding shadow eats the last words of the line.

Minute books

A minute book is evidence of what a company decided, and its physical state is part of that. Photograph the cover and the spine. Note pages that are pasted in, pages with staple holes and nothing attached, and gaps in the numbering. Convert the typed minutes for search, and treat the signatures as images.

Treatises and old law books

Note the title page, the edition, the year and any supplement on the first sheet of your scan set, so every excerpt can be traced to the right edition. Volumes from earlier centuries bring typefaces that recognition may handle poorly, with a long s that looks like an f and letters joined into ligatures. The post on OCR for old documents covers those.

A worked example

Suppose the fictional Tamsin Ridge Water Company is being sold, and the buyer's counsel asks for the board minutes from 1988 to 1996. They are in Minute Book 3, a post-bound volume of 412 numbered pages. A paralegal checks that the posts can be unscrewed and gets the client's written consent to do so. She scans the loose leaves on a flatbed in grayscale, one page per image, including nine blank pages and four resolutions stapled to the minutes of particular meetings, each scanned in place and then lifted. She names the files by printed page number, builds five PDFs of about 85 pages each, converts them on a plan whose per-document page limit covers that size, and exports a searchable PDF for the data room and Word for the lawyer drafting the disclosure schedule. Her page list shows 412 numbered pages, four inserts and one page, 233, numbered twice in the original.

Checking a converted chapter against the printed volume

Check a converted chapter in two passes: completeness first, then accuracy. Completeness asks whether every printed page is present once and in order. Accuracy asks whether the words you will quote or cite match the print. A missing page is the more serious failure and the cheaper one to catch.

Pass one: is every page there?

1. Write down the first and last printed page numbers of the chapter and work out how many pages that makes.
2. Count the images in your file for that range. Add any unnumbered plates or inserts you scanned.
3. If the counts differ, find the break. Look at the page numbers at the quarter points, then narrow down.
4. Check the left-right pattern through the range, if your volume has one.
5. Read the last line of each page and the first line of the next for a handful of page breaks. The sentence should continue.
6. Compare the chapter's headings in your file against the table of contents.

A page list is the record of this pass. One row per image is enough:

Pass two: are the words right?

You cannot proofread a whole book, and you do not need to. Spend the effort where an error would matter:

• Every passage you will quote. Read it against the print, word by word, including punctuation.
• Every pinpoint citation. Open the page image and confirm the passage is on the page you are citing.
• Numbers. Section numbers, dates, amounts, and citations to other volumes.
• Proper names. Parties, places and authors, which no dictionary corrects.
• The inner margin. The first words of lines on right-hand pages and the last words of lines on left-hand pages.
• Footnotes on which the passage relies.

For the wider method, including how to sample a long document, see the post on why OCR output is full of errors.

When the check fails

Rescan the page; do not patch the text from memory or context. If a page is missing from the scan, go back to the volume. If the volume itself lacks the page, record that in the page list in plain words, such as "page 88 absent from the bound volume as examined", with the date you looked. Keep the page list with the scan and the exports. Someone who finds a gap in the text a year from now can then tell whether it comes from the book or from the scanning.