Use enough resolution
For normal printed documents, about 300 DPI is a useful target. Very low-resolution screenshots can blur similar characters such as 0 and O, 1 and I, or 5 and S. Extremely large images do not automatically improve recognition; clarity and focus matter more than raw pixel count.
Keep pages straight and evenly lit
Deskew rotated pages before OCR. When photographing a document, hold the camera parallel to the page and use diffuse light from both sides. Avoid fingers, folds, glare, and phone shadows over text.
Choose the correct output mode
| Source | Recommended mode |
|---|---|
| Letters, reports, contracts, articles | Document or Docs text output |
| Invoices, statements, schedules, price lists | Sheets or table output |
| Mixed page with paragraphs and one table | Start with document mode; extract the table separately if structure matters |
Improve contrast without destroying characters
Text should stand out clearly from the background. Gentle contrast adjustment can help faded pages, but aggressive thresholding may erase punctuation or thin strokes. Keep an unedited copy and compare the result.
Reduce the page complexity
Crop unrelated borders, scanner controls, neighboring pages, and background objects. If a photo contains several receipts, process each receipt separately. For dense multi-column documents, splitting pages into logical regions may improve reading order.
Handle tables carefully
Table extraction works best when rows align consistently and cells contain legible values. Borders help but are not always required. Watch for merged headers, wrapped descriptions, handwritten annotations, and totals printed outside the main grid.
Review high-risk characters and fields
- Account, invoice, policy, and reference numbers
- Decimal points, commas, minus signs, and parentheses
- Email addresses and URLs
- Names with uncommon spelling
- Dates with ambiguous day/month order
- Currency symbols and tax rates