Image Translator

Read the text in a picture and translate it. Upload the image, and the text is detected, recognised and translated. No account, and the upload is deleted after processing.

Two separate things are happening

This is not a file conversion, and understanding why matters when the result disappoints.

First the image is read. Optical character recognition finds the regions that look like writing and works out which characters they are, turning pixels into text. Then that text is translated.

Both steps can go wrong, and they fail differently. If the recognition misreads a word, the translation faithfully translates the wrong word and the output is confident nonsense. If recognition was perfect but the phrase is idiomatic, you get an accurate but wooden rendering.

Telling them apart is easy once you know to look. Compare the detected text against what is actually in the picture. If it does not match, the problem is the photograph, and retaking it fixes everything downstream. If it matches but the translation reads oddly, that is machine translation doing its best with an awkward phrase, and no amount of re-photographing will help.

What reads well

Well: printed text in ordinary typefaces, dark on light, photographed square-on in good light. Signs, menus, packaging, forms, book pages, screenshots and printed letters are all comfortable cases.

Poorly: handwriting, which remains hard even for the best systems and is close to hopeless when cursive. Decorative and display fonts, where letters are stylised past the point of being recognisable individually. Very low resolution, where characters are only a few pixels tall. Text at a steep angle, which distorts letterforms. Text over photographs or patterns, where the background competes. White text on dark can read fine or not at all depending on contrast.

Vertical text, and scripts set right to left, are variable — detection often succeeds while the reading order comes out scrambled.

How to photograph text so it reads

Nearly every poor result traces back to the photograph, and the fixes take seconds.

Square on. Hold the camera parallel to the page rather than at an angle. Perspective stretches letters and recognition degrades quickly with it.

Fill the frame. Get close enough that the text occupies most of the picture. Cropping afterwards does not help, because the detail was never captured.

Even light, no glare. A bright reflection erases whatever is under it. Shift your angle slightly, or shade the page. Avoid shadows across the text, including your own.

Hold still. Motion blur destroys letterforms. Brace the phone against something if the light is poor.

One block at a time. A page with several columns, captions and a photograph gives the detector competing regions and the reading order often comes out jumbled. Photograph each block separately.

How to translate text in an image

1. Upload the image, or drag it onto the upload area.
2. Wait while the text is detected and recognised.
3. Read the translated result.

There is nothing to configure — no source or target language to pick.

What it will not do

Rebuild the layout. You get text, not a translated copy of the original with the same design. Columns, tables and captions flatten into a reading order that is usually sensible and sometimes not.

Replace the text in the picture. The image is not edited, so nothing is painted back over the original.

Read handwriting reliably. Printed text is the realistic case.

Produce a legally or medically dependable translation. Machine translation is excellent for understanding and unsuitable for anything consequential.

A word about what you upload

The obvious use for a tool like this is photographing a document you cannot read, and that category includes passports, residence permits, bank letters, contracts and medical results.

Your image is sent to a server to be processed and is removed afterwards. Even so, for identity documents and anything with a legal or medical consequence, a tool like this is the wrong instrument — both because of what you are transmitting and because the output is not dependable enough to act on. Use it to understand roughly what a document says, then take anything that matters to a human translator.

Why a result can be confidently wrong

This is the failure worth recognising, because it does not look like a failure.

Recognition rarely says "I could not read this". It returns its best guess, and the translation step then treats that guess as fact. A misread digit becomes a different number. A misread name becomes a different word entirely and may be translated as though it were vocabulary rather than a name.

So when the result matters at all, check the recognised text against the picture before trusting the translation. It takes a moment and catches the errors that would otherwise pass silently.

FAQs