Recognise text in scanned PDF pages

Key takeaways

• Text recognition runs locally, without uploading and without credits.
• Docsi uses Tesseract with the official models for German and English.
• Straight, high-contrast scans at around 300 dpi give the best results.
• Always check amounts, dates, names and reference numbers against the image.
• Only accept reviewed text as the basis for further tasks.

Image page or text PDF?

For a normal text PDF, Docsi can read text directly. A scan has no text layer. When you use the PDF as a source for a task, Docsi automatically recognises scanned pages that have not yet been read. The bundled German and English models work together on your device.

A straight, sharp page with good contrast is a better starting point than a blurred photograph. Handwriting, formulas and complex tables are not reliably reconstructed.

Start a task and compare the result

1. Add the scanned PDF as a source. If you first need only the text, use a request such as “Convert this PDF to Word”.
2. Wait for the “Recognize text” step. Docsi reads missing scanned pages in sequence and automatically uses the recognised text for the task.
3. Compare names, numbers, amounts, dates and table rows in the result against the original pages. Pay attention to warnings about scanned sources.
4. Correct errors in the document you continue working with and check important source references. Save a separate output file; the original PDF remains intact.

Recognition is not an accuracy guarantee

Recognised text can contain wrong letters or transposed numbers. Check important details even when the result reads fluently. Docsi keeps recognised page text in the local source, so reusing that source does not replace your own review.

Recognition itself does not upload the scan. What happens afterwards depends on your request: conversion alone stays local; a cloud AI task can then send the recognised text it needs to the provider.

How local text recognition works

Docsi uses Tesseract.js 7 with the official tessdata_fast models for German and English, using LSTM recognition only. The page shown is prepared as an image of at most four million pixels and 2,400 pixels per side. Preparation ends after 30 seconds at most, recognition after 60 seconds; if you cancel or change the page, recognition stops.

Everything runs on your device: neither the PDF nor the page image nor the recognised text is uploaded, and Docsi keeps no database of scanned content. Only when you accept the reviewed text does it replace that one page’s text as a source. The original file and its checksum stay unchanged.

Which scans are recognised well

The Tesseract documentation names the main factors: sufficient resolution (around 300 dpi is a good guide for normal text), straight pages, clear contrast and little noise. A sheet photographed at an angle with shadows produces considerably more errors than a flat scan.

Multi-column layouts, ruled tables, formulas, stamps and handwriting remain difficult. Docsi does not reconstruct page layout: tables arrive as lines of text, formulas as character strings. If you want to analyse a table from a scan, check every row or type critical values into a table of your own.

A review routine for recognised text

First check everything a spell checker will not catch: amounts, dates, names, reference numbers and units. Typical confusions are 0 and O, 1 and l, 5 and S, and decimal comma versus point. Then read sentence starts and paragraph breaks; on multi-column pages the reading order can jump.

For a long scan, work page by page and accept only reviewed text. Note which pages you have recognised and checked before you start an AI task over the whole document. That way you know later which source markers rest on reviewed text.

Photos and screenshots instead of PDF

Besides scanned PDFs, Docsi also accepts images as a source: PNG, JPEG and WebP, such as a photo of a whiteboard or a screenshot. Docsi scales large photos to a compact copy of at most 2,400 pixels per side and reads them locally with text recognition. For AI models that understand images, the image is passed along as well.

For good results, photograph straight from above, in even light and without reflections. Crop unnecessary margins and prefer two halves of a page over a whole page shot from far away. Screenshots are usually easier to read than photos because they are sharp and undistorted.

Scans containing personal data

Because text recognition runs locally, ID cards, invoices or medical letters do not leave your computer during recognition. It is different if you then start an AI task with that text: the text it needs travels encrypted through the Docsi server to the AI provider. Docsi does not store this content, but it is processed for the task.

For sensitive documents, consider whether you only need the recognised text – then everything stays local – or whether an AI task is necessary. Before an AI task you can also add only the pages you need, or remove personal details from the recognised text first.

A worked scenario

An example: you have a scanned chapter from an older book as a PDF, 18 pages without a text layer. In the preview you start text recognition for page 1, compare the text with the image and correct two swapped digits in a year. Only then do you accept the text and move on to the next page.

After all 18 pages you ask Docsi for a summary with page numbers. Because every page has been checked, you can trust the source markers. For a table on page 12 whose columns shifted during recognition, you type the six values into a small Excel sheet yourself and add it as an extra source.

Key terms explained

OCR

Optical character recognition: recognising text in images and scanned pages.

Text layer

The invisible, selectable text of a PDF. If it is missing, the page consists only of an image.

LSTM

The neural recognition method of Tesseract that Docsi uses exclusively.

dpi

Dots per inch, a measure of scan resolution. Around 300 dpi is a good guide for normal text.

Recognition confidence

An engine estimate between 0 and 100. It is not a probability that the text is correct.