OCR PDF (Make Scanned PDFs Searchable)

About the OCR PDF (Make Scanned PDFs Searchable)

A scanned page or a photo of a document is only a picture: you can’t search it, select a sentence or copy a figure from it. OCR (optical character recognition) reads the letters in the picture and turns them back into text.

This tool runs Tesseract 5, the open-source OCR engine, inside your browser. You get a searchable PDF — your original pages, untouched, with an invisible layer of recognised text on top, so you can search it, select it and copy from it in any PDF reader — plus the text as a .txt file and as hOCR (text with the position of every word). Nothing is uploaded: the files never leave your device.

Common uses

• Making scanned contracts, receipts and old letters searchable before archiving them.
• Copying text out of a photo of a notice, a book page or a printed form instead of retyping it.
• Preparing scans for PDF to Word or PDF to Text, which need a text layer.
• Making documents usable with screen readers and translation tools.

What “searchable PDF” means here

Each page keeps its original appearance — for a PDF, the page itself is not changed at all; for images, a JPEG’s compressed picture data is used as it is (no recompression) and PNGs are stored losslessly. Hidden photo details such as the camera and GPS location (EXIF) are not copied into the PDF. The recognised words are added on top as invisible text: text rendering mode 3, “neither fill nor stroke” (ISO 32000-1:2008, §9.3.6). Every word is placed over its picture, so searching highlights the right spot and selecting copies the right words.

The invisible text uses a special font without visible shapes in which each character code is the Unicode character itself. That is why Hindi, Tamil, Chinese and other scripts can be copied and searched even though no font for them is embedded. Arabic and Urdu are stored from left to right, as drawn, which is how PDF readers expect right-to-left text; each reader puts it back into reading order in its own way (see the limitations below). Chinese text is stored without spaces between its characters, as it is written, so a phrase can be found as you would type it.

Getting the best results

• Resolution: 300 DPI is right for normal print. Very small print benefits from 400 DPI; 200 DPI is faster but less accurate.
• Straightening: crooked scans are straightened before recognition, and the text layer follows the original slant, so the page itself still looks exactly as it did.
• Rotation: with auto-rotate on, Tesseract’s orientation detection checks each page and turns sideways or upside-down pages upright — only when it is confident.
• Pages that already have text (from a word processor, or OCR’d before) are skipped by default, so they don’t get a second, duplicate text layer.
• Low-confidence words are marked in yellow in the previews. You can also add them to the PDF as highlight annotations, which any PDF reader can show, hide or delete.

Languages

Text can be recognised in English, Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Urdu, Arabic, and Chinese (Simplified and Traditional), using Tesseract’s standard trained models (the integer versions of its “best” LSTM models from the official tessdata set). Each language has its own recognition data of 1–3 MB, downloaded from this site only when you choose that language.

Many documents mix English into another language — forms, bills and notices in India often do. Tick Also recognise English words and both are read: Tesseract tries the document’s language first and English where that fits better. It downloads the English data too (about 3 MB) and can be a little slower.

Choosing the wrong language gives poor results, so pick the language the document is written in. Arabic and Urdu read right to left: the recognised text and the hOCR keep their reading order.

Limitations

• Handwriting is not recognised reliably — Tesseract is designed for printed text.
• Only the 13 languages in the language menu can be recognised; text in any other language or script comes out wrong.
• Chinese (or any text) set in vertical columns is not read correctly — only horizontal lines are.
• Urdu printed in the Nastaliq style, the usual one for Urdu, is often recognised much less accurately than Naskh-style print.
• In the searchable PDF, copying or searching Arabic and Urdu depends on the PDF reader: some readers change the order of the words on a line, or split it. The .txt and hOCR files don’t depend on the reader.
• Tables are recognised as lines of text, not as table cells. You can try PDF to Excel on the searchable PDF to rebuild them.
• Encrypted PDFs (with a password or editing restrictions) can be read but not rewritten here, so their searchable copy is rebuilt from images of the pages: text and drawings become pictures, and the copy has no password or restrictions.
• Recognition usually takes a few seconds per page on a computer and longer on a phone. Very long documents are best split into parts, especially on phones.
• The first run needs an internet connection to download the OCR engine and language data (about 5–7 MB, depending on the language, and 3 MB more with English words); your browser normally keeps them in its cache for later runs.

Privacy

Everything happens in your browser. What you enter or open here is not uploaded or stored by MySmartCoPilot. The first time you run OCR, the recognition engine and the data for the language you choose (about 5–7 MB, and 3 MB more to also recognise English) are downloaded from this site. Your files are never uploaded.