Image to text · JPG JPG to text: read the words in a photo or scan
Phone photos of documents, scanned pages and screenshots saved as JPG: add them, pick the language and get the text to copy or save. Small photos are enlarged and dark ones inverted before reading, all on your device.
- Photos stay on your device
- English and Chinese
- Small photos enlarged first
Convert JPG to text
Add images, choose the language and click Recognise text. The first run downloads the engine and the language data once; after that the browser cache is used.
No images yet. Choose, drop or paste images, or try the sample image.
English 2.8 MB, Chinese about 1.6 MB, downloaded once. Mixed text needs both.
Use “Scattered text” for posters, memes and app screenshots with text in many places.
What the tidy-up does
Tesseract puts a space between Chinese characters and keeps every line break of the picture. The tool removes the spaces inside Chinese, rejoins words split with a hyphen and, if you ask, joins lines into paragraphs.
Invoice No. 2026-118
Thank you for your pur-
chase. Payment is due
within 14 days.
发 票 号 码 2026-118
感 谢 您 的 购 买 , 请 在
14 天 内 付 款 。Invoice No. 2026-118 Thank you for your purchase. Payment is due within 14 days.
发票号码 2026-118 感谢您的购买,请在 14 天内付款。The tidy-up only touches spacing and line breaks. It never changes the letters Tesseract read, so a misread character stays misread; check names, numbers and amounts by eye.
Version v1.0Updated Update history
How far to trust the text
OCR guesses each character from its shape. On clean printed text it is very accurate; on difficult pictures it makes mistakes that look plausible.
Confidence
Tesseract’s own average certainty for the words it read, 0–100. Above 85 the text is usually right apart from the odd character; between 65 and 85 expect several errors; below 65 read it against the picture line by line.
What works well
Screenshots, exported slides, printed documents and receipts photographed straight on, with text at least about 20 pixels tall. Small images are enlarged automatically, which helps.
What does not
Handwriting, decorative or very thin fonts, text on busy photos, curved or strongly tilted lines, and complex layouts such as tables, forms and multi-column magazines, where the reading order can get mixed up.
Chinese and mixed text
Choose the language of the picture. “Simplified Chinese + English” handles Chinese text with English words and numbers; English only will turn Chinese into nonsense, and Chinese only reads English less well.
For a PDF that already contains text (you can select words in a viewer), use PDF to text instead: it reads the real text with no recognition errors. For a scanned PDF, turn the pages into images with PDF to image first, then add them here.
Turn an image into text in four steps
The defaults (automatic layout, small images enlarged) suit most screenshots and scans.
-
1
Add images
Choose files, drop them on the box or paste a screenshot with Ctrl+V or Cmd+V. You can add several and they are read one after another.
-
2
Pick the language
English, Simplified or Traditional Chinese, or Chinese with English. The language data downloads the first time you use it.
-
3
Recognise
Click Recognise text. Each image shows its progress and, when done, a confidence score.
-
4
Check and save
Read the text against the picture, especially numbers and names, then copy it or download a .txt file.
What happens to the picture before and after OCR
The engine is Tesseract 5 compiled to WebAssembly (tesseract.js), using its LSTM neural-network recogniser. The tool prepares each image the way Tesseract’s documentation recommends, then tidies the output.
- Enlarge: images whose longer side is under 1,200 pixels are scaled up 2×, so letters reach the 20–30 pixel height Tesseract reads best (capped at 16 megapixels)
- Light background: transparent areas count as white paper, and dark-mode screenshots (average grey below 110 of 255) are inverted to dark text on light
- Grey and black-and-white: Tesseract converts to grey (0.299 × red + 0.587 × green + 0.114 × blue, ITU-R BT.601) and picks the black / white threshold with Otsu’s method itself. The optional “black and white first” step does the same in advance; it can help faint text on coloured paper, but in our tests it made very small text worse, so it is off by default
- Recognise: Tesseract finds blocks and lines, then reads each line with the LSTM model for the chosen language; slightly tilted pages are straightened first
- Tidy: spaces between Chinese characters and before Chinese punctuation are removed; with “Join lines” the lines of a paragraph are joined (no space between Chinese, a space between English words, hyphenated words rejoined)
| Language | Model | Download (gzip) | Good for |
|---|---|---|---|
| English | tessdata_best, integer | 2.8 MB | English and other Latin-script text with English words |
| Simplified Chinese | tessdata_best, integer | 1.6 MB | Mainland China screenshots, documents, receipts |
| Traditional Chinese | tessdata_best, integer | 1.6 MB | Taiwan, Hong Kong text |
A small phone screenshot
A 600 × 300 crop of a chat with 14-pixel text, English, default settings.
| Step | What happens | Result |
|---|---|---|
| 1. Enlarge | Longer side 600 < 1,200, so scale 2× | 1,200 × 600 px, text about 28 px tall |
| 2. Check the background | Average grey is well above 110, so no inversion | Dark text on light, as Tesseract expects |
| 3. Recognise | English LSTM model, automatic layout | Text plus a confidence score, e.g. 90 (high) |
| 4. Tidy | Spaces and line breaks cleaned | Ready to copy or save as .txt |
In our test, a crop of 10-pixel English text came back 94.5% character-for-character without enlarging and 97.3% with it; a 13-pixel Chinese crop photographed at a slight angle went from 58% to 81%.
Where image to text goes wrong
Most bad results come from the picture or the language setting, not the engine.
Wrong language
English-only recognition turns Chinese characters into random Latin letters. Pick the language of the picture, or Chinese + English for mixed text.
Blurry or tiny text
Text a few pixels tall cannot be read reliably. Zoom in before you take the screenshot, or photograph the page closer and straight on.
Using OCR on a text PDF
If you can select the words in a PDF, the text is already there. PDF to text gets it exactly; OCR would only add errors.
Trusting numbers blindly
OCR can read 8 as B or 1 as l and still sound confident. Check amounts, dates, phone numbers and codes against the image.
The same privacy-first approach in our analytics
This tool reads your images on your device. TapCub applies the same idea to analytics: collect what you need, keep personal data out.
Cookieless visitor ID
Visitors are told apart by a keyed hash of the site, a salt, the IP address and the browser type. Nothing is written to the browser and the IP is never stored.
Learn moreCompliance centre
Choose balanced or strict, switch on IP truncation for some regions only, set the geography level and retention.
Learn moreSafeguards
Encryption in transit, sensitive fields encrypted at rest, masking by role, export and deletion of one person's data.
Learn moreImage to text questions
Still have a question?
Chat with the team behind TapCub — we usually reply within a few hours on business days.
How do I convert a JPG photo of a document to text?
Photograph the page straight on in good light so it fills the frame, add the JPG here, choose the language and click Recognise text. If shadows or paper texture get in the way, try “Convert to black and white first”.
Does JPG quality affect the result?
Yes. Heavy JPEG compression leaves smudges around letters that OCR can misread. Use the original photo rather than one sent through a chat app, which re-compresses it.
Why does my photo come out as gibberish?
Usually the language is wrong, the text is too small, or the page is photographed at an angle. Choose the right language, crop closer and shoot straight on.
Can I convert JPG to text on iPhone or Android?
Yes. Open this page in the phone browser, tap the box and pick photos from your gallery. Everything runs on the phone, though a large batch is slower than on a computer.
Update history
New features, fixes and data refreshes for this tool, newest first.
-
· v1.0
LaunchedFirst release.
Missing something or found an error? Tell us — requests that come up often go into the next update. All tool updates
Analytics that keep personal data out
TapCub counts visits without cookies and never stores IP addresses, with a compliance centre for stricter regions.