1. Resolution: the single biggest lever

OCR engines read the pixel shape of each letter. Fewer pixels per letter = more guessing = more errors. The rough rule: characters should be at least 20 pixels tall. If your image has text that looks small on screen, zoom into the source before capturing.

Scaling up a small image in an image editor does nothelp — you're just enlarging pixels, not adding information. Always capture at the right size originally.

Concrete tips:

• On a phone, move closer or zoom with the camera lens, not the photo app.
• For a screenshot, zoom the web page in-browser first (Ctrl + +), then capture.
• For a scan, use at least 300 DPI. 600 DPI for small text.

2. Contrast: dark text on light background

OCR models are overwhelmingly trained on black-on-white text. The further your input strays from that, the worse results get. A few specifics:

• Dark mode screenshots(white text on dark background): use the "Invert colors" toggle in extractmytext.app, then re-run. Accuracy usually jumps 10–30 points.
• Faded printed text:enable "Increase contrast" to push the letters back to pure black.
• Colored backgrounds:enable "Grayscale" first — this removes color noise that the engine might interpret as character strokes.

The three preprocessing toggles in extractmytext.app (Grayscale, Invert, Increase Contrast) stack. The most common magic combination for iffy images is Grayscale + Increase Contrast.

3. Crop aggressively

Any non-text content in the image — photos, logos, UI chrome, decorative borders — gives the OCR engine extra work and extra chances to make mistakes. Before you upload, crop tight around the text. You can do this on every OS:

• Mac: open in Preview, click the markup toolbar, choose the rectangle select tool, then Tools → Crop.
• Windows: open in Photos, click Edit Image → Crop.
• iOS/Android: built-in Photos app → Edit → Crop.

4. Pick the correct language

Running French text through an English model works badly — the engine tries to force accented characters into unaccented equivalents and mangles words. Always match the language to your text. In extractmytext.app, 12 languages are available in the dropdown.

For mixed-language images (e.g., English article with occasional French phrases), use the dominant language. Minor second-language content is usually handled OK.

For CJK (Chinese, Japanese, Korean), the correct language is essential — they use entirely different character sets and English mode will produce gibberish.

5. Deskew and deblur

If you photographed text at an angle, the letters are slanted — and Tesseract really does work better on straight text. Any phone Photos app has a "Straighten" tool in the edit panel. Use it.

Blurry text (camera shake, out-of-focus) can't be reliably recovered — retake the photo with more light and a steadier hand. A tripod or just propping your phone against a book helps a lot.

6. Lighting for photos of paper

When photographing physical documents:

• Use diffuse natural light (near a window on a cloudy day is ideal) rather than direct overhead light.
• Avoid shadows from your phone or body falling on the page.
• If you must use a lamp, angle it from the side to reduce glare.
• Turn off the flash — it creates hotspots.

7. Font matters more than you'd think

Standard body fonts (Times, Arial, Helvetica, Georgia, Verdana) are essentially perfect for OCR. Problems arise with:

• Decorative/display fonts: stylized wedding invitations, brand logos, fantasy-novel chapter headings. Tesseract was not trained on these.
• All-caps condensed fonts:the letters touch each other and the engine can't segment them.
• Script/cursive fonts: connected letters are essentially handwriting.

If your source has unusual fonts, there's no clean workaround other than accepting lower accuracy or using a cloud OCR service with stronger models.

What about handwriting?

Tesseract (and by extension extractmytext.app) was not trained for handwriting. Neat block-letter printing works reasonably well; cursive or fast-scrawled notes will not. If handwriting is critical:

• Microsoft OneNote has built-in handwriting OCR ("Copy Text from Picture") that works surprisingly well.
• Google Keep on Android offers similar handwriting recognition.
• Apple's Notes app on iOS 16+ has Live Text for handwriting in photos.

A troubleshooting checklist

When accuracy is below what you need, run through these in order:

1. Is the text clearly readable to you? If not, OCR won't do better. Recapture the image.
2. Are the characters at least 20px tall? If no, capture at higher resolution.
3. Is the language setting correct?
4. Is the image cropped to just the text?
5. Is the text on a plain, high-contrast background? If not, enable Grayscale + Increase Contrast.
6. Is the text white-on-dark? Enable Invert Colors.
7. Is the image straight? If not, rotate before uploading.
8. Still bad? Font or handwriting is likely the issue — no free tool will help much.

Understanding the confidence score

extractmytext.app shows a confidence percentage alongside your result. Rough thresholds:

• 90%+: probably flawless. A quick glance-check is enough.
• 75–89%: mostly correct. Proofread for occasional letter swaps (o/0, l/1, rn/m).
• 60–74%: serious errors present. Try preprocessing, a better crop, or recapture.
• Below 60%: the engine is mostly guessing. Fix the input image before trusting any output.