How many dpi does book scanning really need?

27 Sep 2026

A dozen or so books in my collection have incomplete digital copies, and I plan to have a vendor rescan them on a V-shaped book scanner: about 9,500 pages. The quote is tiered by resolution, and 300 dpi and 600 dpi are far apart. The scans will go through OCR and a single-character classifier, so the question can be asked precisely: does 600 dpi make recognition better?

First, check the sample is real resolution §

The vendor sent a sample with the same page at 300 and 600 dpi. Working back from a page width of about 158 mm, the two images are 1,872 and 3,761 pixels wide, which is 301 and 605 dpi. That is genuine optical resolution, not upscaling.

I cropped the same paragraph of body text and some small table text and compared them side by side at the same display size. At 600 dpi the stroke edges are cleaner and thin horizontals more solid; but at 300 dpi every character is complete and clear, and neither small type nor decimal points are merged or broken.

How many pixels does recognition need? §

Body text at around 10.5 pt is about 3.7 mm tall. At 300 dpi a character is about 44 pixels tall; at 600 dpi about 88.

My character classifier takes 64×64 input, and the OCR engine rescales text lines to a fixed height internally. A 44-pixel character gets scaled up a little; an 88-pixel one gets scaled down. For both models, pixels beyond a certain minimum carry no new information.

The most convincing comparison: what I already have §

The current copies are PDFs made with a phone scanning app. The contents pages are images of 1,714×2,446 to 2,230×3,000 pixels, only two or three hundred KB of JPEG per page, which works out to an effective 230 to 300 dpi, with the perspective distortion, blur and uneven light of hand-held capture on top.

On exactly that material, the character classifier scored above 99.9% on human-verified crops. The only real trouble came from two pages with heavy motion blur, and that was camera shake, not a lack of resolution.

A V-shaped book scanner lays the book flat, with no blur and even lighting. At 300 dpi it is already better than what I have in every respect.

Conclusion, and how I’d negotiate §

Body pages at 300 dpi, colour or 8-bit greyscale, JPEG quality 90 or above, or TIFF.
Write no sharpening, no denoising, no automatic enhancement into the contract. Vendors like to sell “clarity enhancement”, and that kind of processing can easily erase the fine strokes of rare characters as noise.
If you want insurance, have only the contents pages and the editorial notes scanned at 600 dpi. That’s about 8 pages per book, a hundred-odd pages in total, 1.4% of the job, and the hardest rare characters happen to be concentrated in the contents.
Before signing, ask the vendor to scan one page from the books you actually need scanned at 300, 400 and 600 dpi, and run it through your own pipeline. A few calls will show the difference more reliably than any sample.

One cost that is easy to overlook is storage. At 600 dpi colour TIFF, 9,500 pages come to roughly 600 GB; high-quality 300 dpi JPEG takes 5 to 10 GB. The delivery format belongs in the contract too.