Start with the clearest available original
A first-generation photo or scan usually contains more useful detail than a screenshot of a screenshot. Repeated resizing and JPG compression can soften character edges and introduce blocks around punctuation.
Use a source where letters remain distinct at normal zoom. Very small text should be photographed or scanned again when possible.
Keep the camera parallel to the page
Perspective makes one side of a line smaller than the other. Hold the camera directly above the page, keep all corners visible, and avoid a steep angle.
Rotate sideways text before extraction. Slight skew may work, but horizontal lines usually keep a more natural reading order.
Use even light and control reflections
Strong shadows can hide thin strokes. Glossy packaging can create bright reflections that erase letters. Use soft, even light and move the camera until the full text is visible.
Crop unrelated content
Remove empty margins, decorative graphics, neighboring labels, and background objects. A smaller text region gives the engine less unrelated detail to classify.
The image text enhancer includes crop, rotation, grayscale, contrast, threshold, and sharpening controls.
Adjust contrast without destroying thin characters
Moderate contrast can separate dark letters from a pale background. Extreme contrast or threshold settings may remove commas, decimal points, accents, and narrow strokes.
Compare the original and adjusted versions. Use the version that keeps complete character shapes rather than the one that simply looks darkest.
Select the language used by most of the text
The chosen model affects alphabets, diacritics, and likely word patterns. Use the source language rather than the language you want for the final output.
For mixed-language pages, extract separate regions when possible. Read the multilingual image text guide for supported models and limits.
Simplify complex reading order
Sidebars, captions, multiple columns, vertical labels, and footnotes can be returned in an unexpected sequence. Crop each logical block separately when order matters.
Use Extract Table from Image when row and column relationships are more important than plain paragraph text.
Review similar characters and critical values
Check O and 0, I and 1, S and 5, B and 8, commas and periods, currency symbols, and accented letters. Verify every name, date, amount, code, URL, and reference number.
Use confidence as a warning signal, not proof
A low score suggests that the whole image needs attention. A high score can still hide one critical error. Compare the complete result with the source before publishing or using it in a decision.
Follow a final extraction checklist
- Use the clearest available original.
- Keep the page flat and text horizontal.
- Crop unrelated graphics and margins.
- Use even light without glare.
- Select the matching source language.
- Compare original and adjusted versions.
- Check punctuation and similar shapes.
- Verify names, dates, amounts, and codes.
- Correct the editable result.
- Keep the source image for comparison.
Picture2Txt uses Tesseract.js for character recognition. MDN explains browser image work with the Canvas API.