OCR is the magic that turns a photo of a page into real, selectable text. Here’s how it works, and what to tweak when it isn’t.
What is OCR?
Optical Character Recognition is a family of algorithms that read characters from an image. Modern OCR has three stages: pre-processing (deskew, binarise, denoise), segmentation (find lines → words → characters), and recognition (classify each glyph).
Classic vs. neural OCR
Classic OCR (Tesseract 3) ran a decision-tree classifier per character. Neural OCR (Tesseract 5 and successors) runs an LSTM over whole lines of text, which handles joined scripts (Arabic, Devanagari) and poor-quality fonts far better.
What actually hurts accuracy
- Low resolution. Below 150 dpi, characters blur into unreadable blobs.
- Skew & curl. A tilt greater than 2° starts to confuse line segmentation.
- Mixed scripts. English + Chinese on the same page requires two language models.
- Decorative fonts. Script and gothic typefaces have no training data; expect 20–30% character error.
- Bleed-through. If the back of the page shows through, binarisation picks up ghosts.
Checklist for great OCR
- Scan at 300 dpi, 8-bit greyscale.
- Enable auto-deskew.
- Choose the correct language in your OCR tool.
- If columns, use the auto layout mode.
- Export to searchable PDF (text overlay) rather than plain text — you preserve the original visual.
Doing it on PDFAPPLE
Our OCR runs automatically when you upload a raster PDF to PDF to Word, PDF to Excel or PDF to PowerPoint. Pro users can trigger OCR as a standalone operation that produces a searchable PDF overlay — ideal for archiving old paper records.
Beyond characters
Document-understanding models now go beyond characters to extract structured fields — invoice totals, dates, addresses — directly from scanned pages. We expect this “OCR ++ LLM” combo to become the standard for financial and legal workflows within three years.
Tags: #OCR #optical character recognition #scanned PDF #Tesseract
Leave a Reply