Image to Text Converter (OCR)

How the text is recognized

ImageCocoon uses Tesseract, the open-source OCR engine originally developed at HP and later by Google, compiled to WebAssembly so it runs in a browser tab. Its current recognizer is a neural network (an LSTM) that reads whole lines at a time, which makes it robust to different fonts and spacing.

Before recognition, small images are enlarged, because Tesseract reads best when capital letters are around 20 to 40 pixels tall; large photos are scaled down so they don’t waste time. After recognition, words the engine is unsure about are filtered out. Those are mostly specks, shadows and the edges of a page in a photo, which would otherwise turn into stray characters. In our tests with screenshots, receipts, small print, dark mode and a tilted photo of a printed page, it got 99 to 100 percent of characters right.

Languages

Twenty-one languages are available: English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Turkish, Indonesian, Vietnamese, Russian, Ukrainian, Arabic, Hindi, Sinhala, Tamil, Chinese (Simplified and Traditional), Japanese and Korean. Each comes as its own file of 0.7 to 3 MB, downloaded only when you choose it and then kept by your browser. The engine itself is about 4 MB, shared by all languages.

Choose the language the text is actually written in: an English model reading French will miss every accent, and one reading Japanese will produce nonsense. Text that mixes languages is best done in two passes, one per language.

Getting the best results

• Screenshots are the easy case: sharp, straight and high-contrast. They usually come out perfect. See screenshot to text.
• Photos of pages work best taken straight on, in even light, with the text filling most of the frame. Shadows, curved pages and steep angles cause most errors. See JPG to text.
• Small text needs enough pixels: a photo of a whole page from far away may have letters only a few pixels tall. Move closer, or crop to the part you need with Crop Image first.
• Layout: “Keep line breaks” mirrors the lines in the image, good for receipts, lists and addresses. Turn it off to get flowing paragraphs, with words hyphenated across lines joined back together.

What OCR can’t do

Tesseract is built for printed text. Handwriting, including neat block capitals, is usually not recognized reliably. Decorative fonts, text on busy photos, very low-resolution images and text set at an angle within the picture are harder too. Tables come out as lines of text rather than rows and columns, and the original fonts and formatting are not kept: you get plain text.

Always read the result before relying on it, especially numbers in invoices, codes and addresses, where a single wrong character matters. The tool tells you when parts of an image were hard to read. For scanned PDF documents, our sister site PDFCocoon has tools built for PDFs.

Your documents stay on your device

People often run OCR on private things: letters, bank statements, IDs, medical forms. Here none of it is uploaded. The engine and language files come from our own domain, are checked against a fingerprint, and then everything happens in your browser. Read more about what to watch for in online image tools.