Optical character recognition looks at an image and decides which letter each shape is. It has no idea what the page meant to say, so a blurred, tilted or shadowed letter is simply guessed. This guide explains the causes of wrong text, what the documentation of OCR engines and vendors says about each, and how to scan so the result comes out right. For the basic steps of running OCR, see How to make a scanned PDF searchable and How to extract text from a scanned document. This page covers the part those guides only touch on: why mistakes happen.
Resolution: too few pixels per letter
A letter that is only a handful of pixels tall has no clear shape left. The Tesseract project, the open-source engine behind many OCR tools, says in its quality guide: "Tesseract works best on images which have a DPI of at least 300 dpi, so it may be beneficial to resize images." Microsoft documents a similar limit for its Read service, which needs text at least 12 pixels high in a 1024 × 768 image, "about 8-point text at 150 dots per inch". That is a floor for small print, not a target.
Archives set their own targets for different reasons. An older guidance page of the US National Archives, which the Archives now marks as superseded, listed 300 to 600 ppi for black-and-white scans of text records and 300 to 400 ppi for grayscale and colour. Treat it as a sign of common practice, not a current rule.
More is not always better. Very high resolutions make large files and slow reading without adding detail the engine can use. In DocPro, pages of a PDF are drawn at up to about 200 dpi before reading, so a PDF scanned at 600 dpi is not read at 600 dpi. Image files are scaled so the long side is at most 4200 pixels in OCR to PDF and 5000 pixels in Extract text, and very small images (under 1400 pixels on the long side) are enlarged first. A practical aim is 300 dpi for normal print, and the largest legible setting your scanner offers for small print, such as receipts or footnotes.
Skew, borders and shadows
The Tesseract guide explains that a skewed page is one "scanned when not straight", and that "the quality of Tesseract's line segmentation reduces significantly if a page is too skewed." Lines that slope are harder to follow, so words are split or merged. The same page warns that dark scanner edges "can be erroneously picked up as extra characters" and that big borders can cause problems.
Phone photos add three more problems: the camera is rarely square to the page, curved paper bends the lines, and a shadow from the phone or a hand falls across the text. Flatten the page, hold the phone directly above it and use light from the side or from above.
Contrast, background and noise
Before reading, an engine turns the image into black and white. Tesseract does this itself, but its guide says "the result can be suboptimal, particularly if the page background is of uneven darkness". It also says some noise cannot be removed in that step, "which can cause accuracy rates to drop." In practice this means:
- Faint pencil, light grey print and thin fax text lose strokes and become unreadable.
- Coloured or patterned backgrounds, watermarks and stamps over text compete with the letters.
- Show-through from the other side of thin paper adds ghost text. Place a black sheet behind the page when scanning.
- Heavy compression leaves blocky noise around letters. Scan or photograph at good quality first and compress later.
Language: a wrong choice gives wrong letters
An engine reads with a language model, which knows the letters and common words of the language. With the wrong one selected, correct shapes turn into the wrong characters, and scripts that run right to left or that join letters can come out as nonsense. If a scan has two languages, both must be selected. Extra languages that are not on the page only slow the reading, so choose exactly the ones present.
DocPro offers five: English, Arabic, Persian, Urdu and Hindi. Other languages, including Latin-script languages with many accents, are not in the list. Accented letters in such text may be misread, so check them by eye. The OCR to PDF result also reports how many words could not be placed in the text layer because no bundled font covers their characters.
Content OCR handles badly
Some content is hard for any print-trained engine:
- Tables. The Tesseract guide notes that the engine "has a problem to recognize text/data from tables" without custom layout analysis. Cells may be read in the wrong order or merged into one line. Treat table output as raw text to rebuild, not as a spreadsheet.
- Handwriting. Print OCR is built for printed type. Cloud services such as Microsoft's Read model have a separate handwriting mode for some languages, but that is not what DocPro runs. DocPro's own guidance is that handwriting is usually read poorly.
- Columns, sidebars and captions. Reading order can jump between columns.
- Small, decorative or condensed type, and long runs of digits where 0 and O or 1 and l look alike.
How to scan so the text comes out right
- Set the scanner to 300 dpi, grayscale or black and white for plain text, and colour only if colour matters. On a phone, use a document-scanner mode and fill the frame with the page.
- Place the page flat and square against the glass or lay it on a plain, dark, matte surface. Press curled paper flat.
- Check that the page is the right way up. Rotate a sideways page with Rotate PDF before reading.
- Trim dark borders with Crop PDF, or for a photo with Edit image, where straightening and brightness can also be adjusted.
- Scan thin or double-sided paper with a dark sheet behind it to remove show-through.
- Keep the original scan. Run OCR on a copy so a better pass is always possible.
Run the OCR in DocPro
- Open OCR to PDF to keep the page image and add a hidden text layer, or Extract text to get plain text.
- Choose the PDF or image. OCR to PDF also accepts several photos and makes one PDF from them.
- Under "Language of the text", select every language on the page and no others.
- For a PDF, type the pages to read in "Pages", such as 1-3, 5, or leave it blank for all pages. A short test on one or two pages saves time on a long document.
- In OCR to PDF, leave "Keep pages that already have selectable text as they are (skip OCR on them)" ticked normally. Untick it to read pages again whose existing text is poor. In Extract text the matching option is "Always use OCR (ignore the text already in the PDF)".
- Press "Make searchable PDF" or "Extract text". When it finishes, note the confidence figure and the word count.
Both tools run on the device. The scan is not uploaded, and the language data is bundled with the page. They are free, with no account.
Check the result with search
The confidence percentage is the engine's own estimate of how sure it was. A low value hints at a poor scan or the wrong language, but a high value does not prove the words are right. A better test is to search.
- Open the output and press Ctrl+F (or Cmd+F). Search for three or four words you can see on the page, including a name and a number. If they are found, the text layer works.
- Search for common wrong forms such as "rn" where "m" should be, or stray symbols, to find bad areas.
- Compare every figure, date, reference number and amount against the original by eye. These are the errors that matter.
- If a whole page fails, rescan that page and read it alone instead of redoing the document.
What DocPro cannot do
DocPro does not clean up a bad image on its own, deskew pages automatically or offer language packs beyond the five listed. It does not rebuild tables, columns or fonts, and it does not read handwriting reliably. If you need to correct recognised text word by word, or need other languages or table recovery, a desktop OCR program or an office suite with editing tools for scanned PDFs fits better. Where a document has legal or financial weight, proofread it against the original regardless of the tool.
Limits and guidance can change, so check the current documentation of any engine or portal before relying on a figure here (figures above were checked in October 2026).
Frequently asked questions
What is the best DPI for OCR?
The Tesseract documentation says its engine works best at 300 dpi or more. For ordinary printed text, 300 dpi is the usual choice. Very small print may need more. Above about 400 dpi the file grows with little gain, and DocPro draws PDF pages at up to about 200 dpi anyway.
Why does OCR confuse 0 and O or 1 and l?
The shapes are close in many fonts, and the engine chooses by context. Digits inside names or codes have little context. Always check numbers and references by eye.
Does a high confidence percentage mean the text is correct?
No. It is the engine's own estimate. Confirm by searching for words you can see on the page and by comparing key figures with the original.
Can OCR read handwriting?
Not reliably in DocPro, which uses a print-oriented engine. Some cloud services offer a separate handwriting model for certain languages, but results vary.
Why did the scan with two languages come out wrong?
If only one language is selected, the other is misread. Select every language that appears on the page in the Language of the text setting.