Image to text converter (OCR)

How it reads a picture

Text recognition, or OCR, finds the lines of text in a picture, splits them into characters and decides which letter each one is, using a model trained on printed text in the languages you pick. This tool uses Tesseract, a long standing open source engine, compiled to run in the browser. The language data, about 3 MB for English, is downloaded once from this site and kept for next time.

Small pictures are enlarged before reading, since the engine works best on large, clear letters, and slightly tilted pictures are straightened. PDF pages are drawn at about 190 DPI and read one by one.

How accurate it is

On clean printed text, very. We read a 10 page report made to look like an office scan, with its pages drawn at 300 DPI on slightly grey paper and stored as JPEG pictures, and compared the result with the text the report was made from:

A real scan with dust, skew and uneven light will score somewhat lower. Results drop further with what the camera gives the engine: phone photos taken at an angle or in poor light, small text, coloured backgrounds, decorative fonts and pages with text in several directions all cost accuracy. Handwriting is not supported at all.

Getting the best result

• Photograph straight on, in good light, with the page filling the frame and no shadow from your phone.
• Pick the right languages. Each language has its own model. A page in French read as English comes out with more mistakes, especially in accented letters; pick up to three when a page mixes them.
• Choose the layout. Automatic finds columns and headings. One block of text suits a single paragraph or a screenshot of a message. Scattered text suits signs, labels and diagrams.
• Crop to the text with crop image when a photo shows much more than the page.

Searchable PDFs

With Also make a searchable PDF switched on, you also get a PDF that looks exactly like your scan, with the recognised text laid invisibly over each word. You can search it, select and copy from it, and the original picture stays untouched. It is the usual way to make a pile of scanned documents searchable. To rebuild a scan as an editable document with paragraphs and tables instead, use PDF to Word.

If something goes wrong

• It found no text. The picture may be too blurry or too small, or set in a language you did not pick. Try a sharper photo, the right language, or a different layout.
• The text comes out in the wrong order. Columns and boxes confuse the reading order. Try One block of text, or crop to one column at a time.
• The PDF already has text you can select. Then it needs no OCR: PDF to text copies the real text out faster and without errors.

How we tested. A 10 page report of public domain text, drawn as a 300 DPI colour scan on grey paper and saved as JPEG pages, read with this tool's settings in Chrome. Each page's result was compared with the source text character by character, ignoring line breaks and spacing.