Images

Image to Text Guide: Extract, Review and Save Printed Words

·

Image to Text turns a picture of printed words into text you can edit, copy or download. Choose one still JPG, PNG or WebP, select the printed language and run recognition. Review the output against the picture before using it.

The useful result is a passage you have checked, not just a completed OCR run. This guide explains the ten language options, file and pixel bounds, background and quality problems, and which actions preserve or replace your edits.

Image to text

Turn one picture into editable text

Choose an image containing printed text, select its Printed language, and press Recognize text. ToolMellow reads the picture in your browser and places its result in Recognized text. Correct that result in the text box, then use Copy or Download text. The download is a UTF-8 plain-text file named recognized.txt.

Start with the clearest original picture available. A screenshot of an article, a photograph of a printed notice or a scan saved as an image can be suitable. OCR identifies characters from pixels; it does not recover a hidden, guaranteed-correct original. Compare the output with the picture before you reuse it, especially when a small error changes a name, date or amount.

  • Choose one still JPG, PNG or WebP; this workspace has no batch input.
  • Select the language printed in the image, which may differ from the website language.
  • Save your corrected text before choosing a different file or starting recognition again.

Tesseract.js READMETesseract.js 7.0.0 API

Check the file size and both image dimensions

All three limits apply together: at most 25000000 bytes, 24000000 pixels in total and 12000 pixels along either side. The file-size allowance is decimal 25 MB, not 25 MiB. Width multiplied by height gives the pixel count. For example, 6000 × 4000 pixels equals 24 megapixels and meets the dimension bounds, provided the file is also within the byte limit and decodes successfully.

A 13000 × 1000 image is only 13 megapixels but exceeds the side limit. A highly compressed picture may have a small file size yet too many pixels. The tool checks file contents rather than trusting an extension: renaming a PDF to .png does not convert it. Animated PNG and animated WebP are rejected. Damaged files and images whose dimensions cannot be read safely can also fail within the advertised size limits.

Check the file size and both image dimensions
CheckAccepted inputPractical response
FormatOne still JPEG, PNG or WebPExport a still image in a supported format; a filename change is insufficient.
File sizeNo more than 25000000 bytesReduce unnecessary image area or export an appropriate image size.
Total pixelsWidth × height no more than 24000000Check pixel dimensions as well as the file-size display.
Longer sideNeither side above 12000 pixelsSplit or crop a very long screenshot into readable sections elsewhere.

Tesseract.js README

Match the OCR model to the printed language

The selector offers ten recognition languages, one at a time: English, French, German, Italian, Portuguese, Spanish, Simplified Chinese, Hindi, Arabic and Bengali. Choose the language of the words in the picture. The page initially selects the corresponding available model for its interface language; the Indonesian and Urdu interfaces instead start with English because this tool has no Indonesian or Urdu OCR model.

The website language switcher and Printed language perform different jobs. Italian recognition is available without an Italian interface. A Simplified Chinese option does not promise a separate Traditional Chinese model, and an Arabic option is not an Urdu model. The workspace has no automatic language detector or combined-language setting. A mixed-language document may need separate crops or another workflow with the required language support.

Match the OCR model to the printed language
Printed languageModel identifierInterface relationship
English / French / Germaneng / fra / deuEach has a matching interface default.
Italian / Portuguese / Spanishita / por / spaItalian is an OCR option; Portuguese and Spanish also have matching interfaces.
Simplified Chinese / Hindichi_sim / hinChinese and Hindi interfaces start with these respective models.
Arabic / Bengaliara / benArabic and Bengali interfaces start with these respective models.
Indonesian / UrduNo corresponding model hereTheir interfaces start with English; they do not add recognition support.

Tesseract.js 7.0.0 API

Keep track of which action replaces your text

Changing Printed language preserves the current text and clears the previous error. It does not run OCR again or translate what is already in the editor. This lets you keep corrections while preparing a different setting. Pressing Recognize text starts a new run and clears the old text immediately, so copy or download edits you want to retain before doing so.

Choosing a new file also clears the previous result. During recognition, the file chooser, language selector and text editor are disabled. After completion, the editor contains the output from that run until you change it or perform an action that clears it. There is no saved-job history or comparison panel: if several pictures belong together, save each reviewed result and assemble them in your own document.

Tesseract.js 7.0.0 API

Improve the source before asking OCR to read it

Use a sharp image with clear separation between lettering and its background. Photograph a flat page with even lighting and keep the camera approximately square to it. Avoid reflections over words and deep shadows along a fold. Cropping unrelated scenery can help, but leave a small margin around the text rather than clipping accents, descenders or the edge of a line. Use Image Cropper separately if you need that preparation.

Straighten severely skewed text and inspect small letters at a readable zoom. Enlarging an already blurred image cannot recreate lost character detail, and changing a DPI label alone adds no pixels. ToolMellow has no crop, contrast or resolution controls in this OCR panel. Repeating recognition on the same poor source is less useful than obtaining a clearer picture or isolating the relevant passage.

Tesseract: Improving the quality of the output

Check the background and orientation

The image decoder uses its orientation metadata. Before recognition, the tool draws the decoded image onto a white background at the decoded pixel dimensions. Transparent regions therefore become white. Black lettering on transparency can work, but white lettering intended for a dark background may disappear. Export that image with a suitable contrasting background before trying again.

The engine may attempt rotation handling during recognition, but that is not a guarantee for sideways, upside-down or skewed text. The page offers no rotation control or preview of the prepared image. If the output is poor, inspect the source in an image viewer and rotate it deliberately in a separate tool. Review the new recognition result rather than assuming that corrected orientation solves every reading problem.

Tesseract: Improving the quality of the outputTesseract.js 7.0.0 API

Treat the output as text that needs reading-order checks

This workspace returns editable plain text. It does not reproduce page design, font styling, spreadsheet cells, illustrations or a searchable PDF. Line breaks may be useful, but they are recognition output rather than an exact map of the original page. Two columns, a price table or a form can produce words in an unexpected order even when individual characters look right.

Check the relationships between labels and values before pasting the result into another document. For a multi-column page, recognizing separate crops can make review easier. If you need a table, reconstruct and verify its rows and columns explicitly. Download text saves what is in the editor; changing its extension to .docx, .xlsx or .pdf does not create those document formats.

Tesseract: Improving the quality of the outputTesseract.js 7.0.0 API

Review the places where a small error changes the meaning

Read the output alongside the original image. OCR can confuse similar shapes, omit punctuation, change accents or separate characters incorrectly. A word that looks plausible may still be wrong. The page does not display a confidence score, and successful completion does not certify every character. Give extra attention to proper names, product codes, dates, totals and text outside the selected model’s usual script.

Use a fictional receipt reading Invoice 4827 — Total 125.50 as a review exercise: check every digit, the decimal mark and which label belongs to which value. Do not infer a correction from what seems likely when the source is unclear; return to a better image or the original document. Make the corrections in Recognized text before copying or downloading.

Review the places where a small error changes the meaning
InspectPossible problemReview action
Names and accented lettersA plausible word replaces the printed name, or an accent is lost.Compare each name with the original rather than trusting spell-check alone.
Numbers and codesO / 0, I / 1 or a missing character changes a value.Compare the full sequence, including leading zeroes and separators.
Amounts and datesA decimal mark, sign or date separator is missing.Check punctuation together with the surrounding label and context.
Tables and columnsCharacters are readable but associated with the wrong row.Verify reading order and rebuild the intended structure explicitly.

Tesseract: Improving the quality of the output

Export the corrected contents of the editor

Copy sends the current editor contents to your clipboard, including your changes and line breaks. After success, the button reads Copied; editing the text resets that label. If clipboard access fails, the page asks you to select the text to copy it. You can use the editor’s normal selection and your browser or operating system’s copy command, or download the text instead.

Download text creates recognized.txt from the current editor contents using UTF-8. The browser chooses the download location or asks where to save it. This action does not preserve the original image or reconstruct formatting. Copy and Download text are unavailable while recognition runs or when the editor is empty. A successful copy retry clears the old clipboard warning.

Tesseract.js 7.0.0 API

Understand what stays in the browser and what still loads

The OCR workflow processes the chosen image in a browser worker. It does not upload the picture or recognized text to an OCR server. The engine and selected language model load from ToolMellow’s own site. The first run can therefore require asset downloads even though character recognition itself happens locally. Each run uses one selected model rather than downloading every recognition language for the job.

Local OCR does not mean that the website makes no network requests: pages, scripts, fonts, OCR files and normal site analytics can still load. Do not assume the tool will work indefinitely offline or that a model is permanently saved. Copying puts text on your clipboard and downloading saves a separate file under your control. Reloading the page loses the editor state; the tool provides no recovery history or secure-deletion guarantee.

Tesseract.js 7.0.0 local installationTesseract.js 7.0.0 API

Use progress and Cancel without treating them as accuracy measures

During a run, the page shows Working and a progress percentage. Engine and model initialization can keep the displayed value at zero before recognition progresses. That percentage is an activity indicator, not a confidence score or a reliable time estimate. Large images, limited device memory and asset loading can affect how long the job takes; this workspace has no fixed recognition-duration deadline.

Cancel stops the current job and restores the controls. It does not pause the engine for later resumption, keep partial OCR output or restore text cleared when the run began. If you start another run, recognition begins again with the selected file and language. If loading appears stuck, cancel, check that the site’s assets can load, and retry with a readable smaller image if appropriate.

Tesseract.js 7.0.0 APITesseract.js 7.0.0 local installation

Separate an empty reading from a failed job

No text was found means the engine returned no non-whitespace text for that run. Check that the picture contains visible printed words, that the selected language is suitable and that white-on-transparent lettering has not vanished against the white background. An empty result does not establish that the source contains no text; it may be unreadable to this recognition setup.

Unable to finish is broader. It can follow unsupported or damaged input, exceeded limits, a decoder or memory failure, or an engine/model loading problem. The message alone does not identify which cause occurred. Work through the relevant checks below, preserving any useful edited text before changing files or starting again.

Separate an empty reading from a failed job
ObservationCheck firstNext step
Recognize text is unavailableNo file selected, or a run is active.Choose one supported image, or wait for or cancel the active run.
No text was foundBlank image, poor contrast, wrong model or difficult lettering.Inspect the source and language, then try a clearer relevant crop.
Unable to finishFile format, all three limits and asset loading.Export a fresh supported image within the bounds; reload if needed.
Garbled or missing wordsSource quality, printed language and reading order.Review against the image and correct the editor; use another workflow if its script is unsupported.
Clipboard unavailableBrowser clipboard access failed.Select and copy the text manually, download it, or retry Copy.

Tesseract.js 7.0.0 APITesseract: Improving the quality of the output

Recognize the boundaries of this image-text tool

Handwriting, signatures and exact page layout are unsupported. This tool is intended for printed words in its ten available recognition languages. It does not translate text, decode barcodes, interpret diagrams or explain the meaning of a document. A legible handwritten note or an Urdu page still requires an appropriate alternative workflow rather than an assumed match to a nearby language option.

For a PDF, use the separate PDF to Word tool and check that tool’s own limits and output behavior; this image workspace does not accept PDF input. Use Image Cropper to isolate text in a picture, or Word Counter after reviewing the extracted text. Choose the next step according to the result you actually need: a verified text passage, a reconstructed document or some other format.

Tesseract.js READMETesseract FAQ

Frequently asked questions

Which image formats can I use?

Choose one still JPG, PNG or WebP. Animated PNG/WebP and unsupported file contents are rejected. Renaming an extension does not convert an image, and a valid-looking file can still fail if it is damaged or cannot be decoded safely.

What are the size limits?

The file may be at most 25000000 bytes, with at most 24000000 pixels total and neither side above 12000 pixels. All three bounds apply together; fitting within the byte limit alone is insufficient.

Which OCR languages are available?

English, French, German, Italian, Portuguese, Spanish, Simplified Chinese, Hindi, Arabic and Bengali. Select one language at a time to match the printed text. There is no automatic or combined-language option in this workspace.

Can the Urdu or Indonesian interface read Urdu or Indonesian text?

Those interface languages do not add OCR models. Both start with English, and this tool has no Urdu or Indonesian recognition model. Arabic recognition is not an Urdu model. Use a workflow that explicitly supports the printed language.

Does changing Printed language translate the text?

No. It chooses the model for the next recognition run and preserves the existing editor contents. To read the image again, press Recognize text; that new run replaces the old text, so save any corrections first.

Can it recognize handwriting or preserve a table?

Handwriting and exact page layout are unsupported. The output is plain text, and table or column reading order can be wrong. Review and reconstruct any structure you need.

Why is the result empty?

The engine may not find readable text with the chosen image and model. Check the language, contrast, clarity and background. Transparent images are flattened onto white, which can hide light lettering.

Does the percentage show OCR accuracy?

No. It reports progress, with initialization sometimes remaining at zero. The page does not show a confidence score, and completion is not a guarantee that the recognized text is correct.

What happens when I cancel?

The current job stops and the controls become available. There is no pause/resume or partial-result recovery. Text cleared at the start of recognition is not restored.

Are the image and extracted text uploaded?

The inspected OCR workflow processes them in your browser rather than sending them to an OCR server. The page and selected OCR assets still load from the site, and ordinary site requests can occur. Local processing is not a promise of permanent offline operation.

Will the download include my corrections?

Yes. Copy and Download text use the current editor contents. The download is a UTF-8 plain-text file named recognized.txt, without the source picture or reconstructed page formatting.

Can I recover text after selecting another image or reloading?

The tool has no saved-job history. Choosing a new image, starting recognition again or reloading can clear the current text. Copy or download edits you want to keep before taking those actions.

Sources and further reading

Put it into practice.

Related guides