How to OCR a scanned document more accurately
Improve recognition by preparing straight, readable scans and learn which OCR mistakes always require human review.
OCR reads pixels, not meaning
Optical character recognition estimates letters and words from an image. It can make a scan searchable or provide editable text, but it does not understand the document well enough to guarantee every name, number or symbol.
Use OCR output as a draft. Legal clauses, medical details, totals, dates and account numbers require comparison with the source.
Improve the scan
Place the page flat, fill the frame and use even lighting without glare. Dark text on a clean light background is easier to recognize than shadows, patterned paper or a low-contrast photograph.
Rotate the page upright and crop away surrounding objects. Higher resolution helps small print, but an enormous blurry photo is not better than a sharp, sensibly sized scan.
- Keep the camera parallel to the page
- Avoid fingers over text
- Use the correct language model
- Review one page before processing a large set
Expect predictable mistakes
OCR can confuse zero with O, one with l, punctuation with noise and columns with a single line of text. Decorative fonts, handwriting, stamps and curved pages create additional errors.
Tables may lose row and column relationships even when individual words are recognized. Preserve the original page image whenever layout carries meaning.
Searchable PDF and extracted text are different
Extracted text is convenient for copying into another document. A searchable PDF keeps the original page appearance and adds a text layer or related recognized output, depending on the tool.
Choose based on the next task. Do not discard the visual scan after copying text because it remains the evidence needed to correct recognition errors.
Handle sensitive scans carefully
MyFileNest runs supported OCR locally, although the browser may first download an OCR model or language data. Downloading processing code is different from uploading your file.
Use a trusted device, review the final output and remove temporary browser data when working on a shared computer.
What affects accuracy most
| Factor | Good input | Problem input |
|---|---|---|
| Sharpness | In focus, steady | Motion blur, out of focus |
| Resolution | About 300 dpi for normal print | Below 150 dpi, or tiny text in a wide shot |
| Contrast | Black on white | Faded toner, coloured paper, pencil |
| Alignment | Straight, flat pages | Skewed, curled near a spine, folded |
| Language setting | Matches the text | Wrong language model |
| Layout | One column of text | Tables, forms, sidebars, text over images |
Choosing the right MyFileNest OCR tool
Image to Text recognises a single photo or screenshot and shows the words in an editable box that you can copy or download as plain text. It is the best choice when you need the exact wording, for example to quote a paragraph or paste figures into another document.
OCR PDF processes a scanned PDF of up to 20 pages and produces a new PDF that keeps the original page images with an invisible text layer, so you can search the document and find the right page. Both tools support English, Spanish, French and German and run in your browser.
Step-by-step: a reliable OCR routine
- Capture or scan pages straight and evenly lit, ideally at around 300 dpi in grayscale.
- Rotate pages upright with Rotate PDF or Rotate Image.
- Crop away dark borders and desk surfaces, which can be misread as characters.
- Choose the recognition language that matches the document.
- Run recognition on one page first and review the result.
- If accuracy is poor, fix the input rather than rerunning OCR: rescan, increase contrast or retake the photo.
- Process the rest of the document and proofread names, numbers, dates and amounts against the scan.
Worked example: digitising old recipes
Someone wants to type up 15 typewritten family recipes. Photographing them on a kitchen counter under a ceiling light produces shadows across each page and recognition full of errors. They retake the photos by a window in daylight, holding the phone directly above each card, and the recognised text becomes almost entirely correct. Running each photo through Image to Text gives editable text to paste into a document; they then check every quantity, because a misread "1/2" as "112" matters more in a recipe than a typo in the method.
Proofreading checklist
- Numbers: compare every total, date, reference and phone number with the image.
- Look-alike characters: 0 and O, 1 and l and I, 5 and S, rn and m.
- Punctuation: decimal points and commas in figures, and hyphens at line breaks.
- Names and addresses: OCR has no way to know an unusual spelling is correct.
- Missing lines: compare the first and last line of each page with the scan.
OCR for different kinds of documents
Knowing the likely accuracy helps you plan how much proofreading time to allow. A ten-page printed letter may need a quick read-through, while a stack of faded receipts needs every figure checked.
| Document | Typical accuracy | Notes |
|---|---|---|
| Printed letters and reports | High | Clean print in common fonts recognises very well |
| Receipts on thermal paper | Moderate | Faded print and small fonts cause errors; check totals |
| Forms with boxes and lines | Moderate | Lines can be read as characters; text order may jump between fields |
| Books and magazines | High to moderate | Curved pages near the spine and columns reduce accuracy |
| Handwriting | Low | General OCR engines are designed for print; expect to retype |
| Old typewritten pages | Moderate | Uneven ink and faded letters need good contrast |
Language settings and mixed-language documents
OCR engines use language data to decide which characters and words are likely. Choosing English for a French document produces more errors, especially with accented letters and common words. MyFileNest OCR tools offer English, Spanish, French and German; choose the language that most of the document is written in.
For documents that mix languages, such as a bilingual contract, recognise the document with the main language, then pay particular attention to passages in the other language when proofreading. For languages not currently offered, desktop OCR software or a specialist service may give better results.
Privacy and sensitive documents
Scanned documents sent for OCR often contain identity numbers, signatures, medical information and financial details. Cloud OCR services can be very accurate, but they process the document on someone else's servers. Running OCR in the browser keeps the scan on your device, which suits personal and confidential documents.
After OCR, remember that the recognised text makes the document easier to search, and therefore easier for anyone with the file to find sensitive details. Store searchable copies with the same care as the originals, and consider whether a searchable version needs to be shared at all, or whether the original scan is enough for the recipient.
Quick answers
Can OCR read text in photos of signs or screens? Often yes, if the text is sharp and fills a good part of the image. Use Image to Text for single photos and screenshots.
Why is the text order wrong? OCR reads layout regions in an order it estimates. Columns, sidebars and tables can be read in an unexpected sequence; rearrange the text when proofreading.
How long does OCR take? It depends on the device and page count; a few seconds to several seconds per page is common in a browser.
After OCR: making the text useful
Recognised text is a starting point. If you need an editable document, paste the text from Image to Text into a word processor, or convert a searchable PDF with PDF to Word, then fix line breaks, headings and any recognition errors. If you need the text in a spreadsheet, such as figures from a scanned table, use PDF to Excel on the searchable PDF or paste rows into a spreadsheet and split them into columns.
For archives, the searchable PDF is usually the most useful end product, because it keeps the original page image as evidence while making the text findable. Name the searchable copy clearly and replace the image-only scan in your filing system, so there is one authoritative version of each document.
When to choose a different approach
OCR is not always the fastest route. For a short handwritten note, retyping it can be quicker than correcting recognition errors. For a document that exists digitally somewhere, such as a bank statement available in online banking or a letter the sender can email, asking for the digital original avoids OCR entirely and gives perfect text. For very large archives, specialist desktop or enterprise OCR software can process thousands of pages in batches with layout analysis. Browser OCR is ideal for everyday documents: a few letters, a scanned contract, a page of a book or a set of receipts, where privacy and convenience matter more than industrial throughput.
Summary
Capture straight, sharp, evenly lit pages, crop and rotate them, choose the correct language, and test one page before processing everything. Treat OCR output as a draft: proofread names, numbers and dates against the image, and keep the scan alongside the text.