How OCR sees a page
Optical character recognition finds lines of text, segments them into words and characters, and matches shapes against a model of the language. Every stage assumes something: that lines are horizontal, that ink is dark and paper is light, that characters are a few dozen pixels tall, and that the language is the one you said. Break an assumption and accuracy falls off a cliff — which is why the same engine gives 99% on one scan and 70% on another. Bindery runs Tesseract on your device with six language packs; the steps below are about giving it a page it can read.
1. Resolution: aim for 300 DPI
Characters need roughly 20–30 pixels of height to be recognised reliably. On an A4 page with 10-point text that means about 300 DPI. Scanning at 150 DPI halves the pixels and roughly doubles the error rate; 600 DPI adds nothing for text and quadruples the file size. Phone photos work when the page fills the frame; Scan from phone straightens the perspective as you capture.
2. Skew: straighten it first
A page rotated even two degrees on the glass makes lines drift across the row segmenter. Clean scanned PDF estimates and corrects rotations up to about three degrees; a page that is sideways needs Rotate first. Deskew is the single change that most improves recognition on real-world scans.
3. Contrast: black ink, white paper
Grey scans of faded originals confuse the threshold that separates ink from paper. Clean scanned PDF lifts weak contrast so paper reads white and ink reads black, and its despeckle option removes isolated dust pixels that the engine otherwise tries to read as punctuation. For text documents, rescanning in greyscale or black-and-white gives cleaner input than colour.
4. Borders and noise
Dark scanner edges, punch holes and binding shadows confuse layout analysis — the engine may decide the black strip is a column. Crop the frame to the paper on every page with one proportional rectangle. Remove blank pages with Delete pages so the engine does not waste time inventing text on noise.
5. Language
Pick the document’s language. English recognition on a French letter loses accents and mangles words; the engine’s dictionary is part of its accuracy. Bindery offers English, Spanish, French, German, Portuguese and Italian, each loaded once and cached. Mixed-language documents: choose the dominant one and check the rest by eye.
6. Order of operations
- Rotate sideways pages; delete blanks.
- Crop borders.
- Clean: contrast, deskew, despeckle.
- OCR in the right language, as a searchable PDF (the page image is kept; the text layer is added invisibly on top).
- Compress last — Balanced keeps scanned text legible.
Never compress before OCR: compression resamples the images the engine needs to read.
Handwriting, tables and stamps
Printed text is the easy case. Handwriting is largely beyond an offline engine tuned for print; expect fragments, and treat the page image as the record. Tables of digits are recognised character by character with no arithmetic sense, so a 3 read as an 8 goes unnoticed unless you check totals. Rubber stamps, watermarks and highlighter overlap the text and produce noise; if a stamp covers a critical line, that line is the one to verify by eye. Two-column layouts recognise well when the columns are cleanly separated; run Crop to one column at a time if the engine merges them.
Checking the result
Search the finished PDF for a few names and numbers you know are in it. Copy a paragraph into a text editor and read it. Spot-check tables of digits, which are the weakest area of every engine. For documents where accuracy is critical — statements, contracts, evidence — the text layer is a search aid and a starting point, not a substitute for the page image, which is why the searchable-PDF output keeps every page pixel-identical.