OCR a PDF
A free tool that reads the text in scanned PDFs with OCR (optical character recognition) and converts them into searchable PDFs. The text it reads is added to the pages as invisible text, so the pages look the same. Your PDF is processed in your browser and never uploaded. You can read up to 200 pages at a time (30 pages on devices where processing takes longer). You can open PDFs of up to 200 MB per file (100 MB on phones and tablets).
Only when a PDF has pages whose text is an image does the first run download an AI model (about 17 MB) from this site; later runs use the copy stored in your browser. Your PDFs are never uploaded.
Settings
How to use
- 1
Add PDFs
Choose or drop the PDFs to read. You can add several PDFs at once. Your PDFs are never uploaded.
- 2
Choose the document language
Under "Document language", choose the language the PDF is written in. To read only some pages, enter them in "Page range".
- 3
Read and save
Press "Run" to read the pages whose text is an image and create a PDF with a text layer. Save it with "Save", and check the recognized lines with "Check the recognized text".
About OCR for PDFs
Text is read with PaddleOCR's text recognition model (PP-OCRv6, Apache-2.0), which runs inside your browser. Only when a PDF has pages whose text is an image does the first run download the AI model (about 17 MB) from this site; later runs use the copy stored in your browser. On devices that cannot use the GPU (WebGPU), processing runs on the CPU and takes longer than on the GPU. After the first page is read, the tool estimates the total time from how long that page took on your device and shows it. The invisible text uses a font made from Tesseract's GlyphLessFont (Apache-2.0).
Pages that already have text data and pages that already have an OCR text layer are not read and stay as they are. On pages that mix text data and images, only the parts that do not overlap the text data are read. Pages whose text data is garbled (selecting the text gives different characters) do not get a text layer. Choosing Japanese under "Document language" limits recognition to the characters used in Japanese documents, and choosing Korean uses a Korean model (about 13 MB, which cannot read Chinese characters). Press "Check the recognized text" to see the recognized lines outlined on the page, with the lines more likely to contain errors shown in a different color.
Specifications of this tool
- Input
- Output
- Limits
- No limit on the number of files (batch processing). Up to 200 MB per file (100 MB on phones and tablets). Up to 200 pages read at a time (30 pages on devices that process on the CPU)
- Free, no sign-up, no app to install
- Files are never uploaded (processed in your browser)
Frequently asked questions
Is my PDF sent to a server?
No. Opening the PDF, reading the text and saving all happen in the browser on your device. Statistics that help us fix bugs (the page count range, the ranges of how many pages and lines were read, the range of the time per page, error types and so on) are sent to this site's server. They do not include the content of the PDF, the recognized text, file names, passwords or the page ranges you enter. You can turn the statistics off with the "Send processing statistics" switch at the bottom of the page.
Can it read a password-protected PDF or one the author has restricted?
Yes. A password-protected PDF opens once you enter its open password. If the author has restricted changing the content or copying and extraction, the tool does not read the PDF until you enter the permissions password (adding a text layer would make the restricted text extractable). This tool does not guess passwords, so if you do not know the permissions password, ask the person who created the PDF. The original PDF's encryption stays as it is in the saved PDF. If the PDF has a digital signature, the signature is no longer valid in the saved PDF.
Can it read vertical text or handwriting?
It is made for printed text written horizontally. It also reads pages of vertical text, but with more errors than horizontal text, and it can miss punctuation. Handwritten text is not supported.
Can the recognized text contain errors?
Yes. Depending on the text size, the scan resolution, and stains or skew on the paper, some characters can be misread or missed. Numbers spaced widely apart in tables are misread more often than other text. When "Japanese" is chosen as the document language, characters outside the JIS level 1 and 2 kanji sets and emoji can be read as similar-looking characters. Before you use the saved PDF, check the lines more likely to contain errors with "Check the recognized text". This tool cannot correct the recognized text.
Help us improve the service
If a conversion fails, sending us the circumstances helps us track down the cause. Knowing which devices and formats fail most often, and how long processing takes, lets us revise the supported formats and fix the bugs behind them. If you don't mind, leaving this on is a real help.
What is sent
- The tool you used and the settings you chose (output format, quality, etc.)
- Your device specs (GPU, memory, core count)
- The input format with a rough size
- How many files succeeded or failed and how long it took
- The error type and details when something fails
What is never sent
- The files themselves and file names
- The text in PDFs, the text you rewrote or redacted, passwords, font names and the page ranges you enter
- The recognized text and the positions of the recognized lines
- IP addresses
The data goes only to this site's own server, never to a third-party service.
For the full list, including the details each tool sends, see “Processing statistics” in the privacy policy.
Details sent by this tool (counts, times, kinds and yes/no only)
In addition to “What is sent”, this tool sends the following details.
- PDF tools (all)
- The page count range, whether editing is restricted, whether a permissions password was entered, and whether the PDF is signed or encrypted.
- OCR for PDFs
- The numbers of inputs and outputs, of pages read, pages read with no text found, pages with text data that were not read, pages mixing text data and images, pages that already had an OCR text layer, pages with garbled text data, blank pages, pages that could not get a text layer, pages whose orientation was corrected and pages where an unclosed section of the original content was closed, of lines and characters read, of areas and characters not read because they overlapped existing text or annotations, and of passes adding the text layer (all as ranges), the ranges of the share of vertical lines and of lines more likely to contain errors, the ranges of the time per page and of the AI model load time, the AI model type and version and whether it was downloaded, the range of the rendering resolution and whether it was lowered, the range of the number of parts an image was split into to find text, whether the estimated time exceeded 15 minutes, whether the GPU became unavailable, and whether the original protection was reapplied.