OCR text extraction from images and PDFs

Clean printed text comes out almost perfectly; phone photos and faded scans are harder. The engine turns sideways shots upright and reads every line through its own tilted frame, and low-confidence words are highlighted so you can fix them before export.

A neural engine reads English, Russian and the other Latin- and Cyrillic-script languages, Chinese, Japanese, Korean, Arabic, Greek, Hindi and Thai; about a hundred more — Hebrew, Georgian, Vietnamese, Bengali and others — go to Tesseract. Languages are grouped by script in the list; models load on first use and stay in the browser cache.

Yes. Multi-page PDFs are rendered page by page via pdf.js and then passed through OCR. You can also export the result back as a searchable PDF with an invisible text layer, so Ctrl+F works and screen readers can read it.

The regular mode is tuned for printed text: neat handwriting with separate letters is often readable, cursive is not. For handwriting there is “Recognize more precisely”: the page is sent to a recognition server and read by a model built for handwriting. On school notebook pages it read nearly 9 words in 10 correctly; the regular mode, fewer than 1 in 10.

After OCR we group recognized words by their vertical position into rows and split rows into columns based on horizontal gaps. When at least three rows share the same column count we treat it as a table and export it as rows and columns in Excel; otherwise the Excel file has a single column of text.

The queue takes up to 50 files. How many files are recognized per day and how many PDF pages one pass reads is shown in the box above the Extract text button. The per-file size cap adapts to your device (roughly 80 MB on phones up to 1 GB on desktops); PDFs are processed page by page, and a big document gets a page-range picker.