PDF OCR — Make Scans Searchable

What is PDF OCR? 🔍

OCR stands for Optical Character Recognition. It's the process of teaching a computer to read text from an image. When you scan a page or take a photo of a document, the result is a picture — pixels, not letters. A human eye can read those pixels, but a computer just sees a grid of colours. OCR bridges that gap.

A PDF OCR tool does two things: it renders each page as an image, then runs a recognition engine over that image to figure out which pixels form which letters. The result is either plain text you can copy, or a searchable PDF where the recognised text is embedded invisibly behind the original page image.

This tool does both — entirely in your browser, using Tesseract, an open-source OCR engine that's been under development since 1985 and is now maintained by Google and the open-source community. It's the same engine used by many commercial products.

Why you'd make a PDF searchable 🔎

A scanned PDF looks fine — you can read it on screen or print it. But you can't search it, select text from it, or copy a paragraph out of it. The words aren't really there; they're just ink on a page image. Making it searchable unlocks three things:

• Search (Ctrl+F). Jump directly to any word in a 300-page document.
• Copy and paste. Pull a paragraph into an email or a report without retyping it.
• Screen readers. Accessibility tools like NVDA and VoiceOver can read the document aloud — an image-only PDF is unusable for blind users.
• Indexing. Search engines and document management systems can only index what's actually text. A searchable PDF is discoverable; a flat scan isn't.

The three kinds of PDF you'll encounter

How to OCR a PDF — step by step 📋

1. Upload your scanned PDF. Drop it onto the upload box or click to browse. The file stays on your device.
2. Pick the language. Choose the language that the document text is written in. English documents need "English", even if you're on a Spanish interface.
3. Adjust rendering DPI. 200 is a good default. Bump it to 300 for small print or dense text.
4. Enable preprocessing if needed. For faded or tinted scans, turn on Grayscale and Contrast boost.
5. Click Run OCR. On the first run, the engine downloads (~4 MB) and the language model (~2–15 MB depending on language). This is a one-time cost.
6. Watch the progress. Each page is rendered, preprocessed, recognized, and mapped — a few seconds per page on a typical laptop.
7. Review the extracted text. Switch to the "Extracted text" tab to see what was recognised.
8. Download. Get the searchable PDF, the .txt file, or both.

How the searchable PDF is built 🔬

This is the interesting part, and it's where a lot of OCR tools cut corners. Here's what happens under the hood:

1. Render — each PDF page is drawn onto an HTML canvas at the chosen DPI using PDF.js.
2. Preprocess — if enabled, the canvas pixels are converted to grayscale and/or their contrast is stretched.
3. Recognize — the canvas is passed to Tesseract, which returns a list of words, each with a bounding box {x0, y0, x1, y1} in canvas pixels and a confidence score 0–100.
4. Rebuild — pdf-lib creates a new PDF. For each page, the original scan is drawn as a background image, then each recognized word is drawn as invisible text (opacity 0) at the position that corresponds to its bounding box.
5. Scale — the key conversion is PDF_points = canvas_pixels × (page_width_pt / canvas_width_px). The font size for each word is set from the word's bounding box height, so highlights line up with the printed text.

How accurate is OCR? 📊

The biggest single lever on accuracy is DPI. Rendering at 72 DPI (the PDF default) gives Tesseract barely-legible text — it essentially guesses. 180–220 DPI is the sweet spot for most documents. Go higher (300–400) only if you have small print or thin fonts.

What's preserved (and what isn't) ⚠️

✅ What the searchable PDF keeps

• The original scan — visible page image, pixel-for-pixel
• Page dimensions — same size, same orientation
• Invisible text layer — searchable, selectable, copyable
• Per-page text alignment — words sit over the right positions

⚠️ What changes

• The visible layout of the text — this is not a text-extraction layout. The output is the same scan with an invisible text layer. If you want editable text on a new layout, use PDF to Word.
• File size — the scan is re-encoded as a JPEG inside the new PDF, and the invisible text layer adds a few KB per page. Expect roughly the same size or slightly larger.
• Text font — the invisible layer uses a standard Helvetica font for Latin scripts. Non-Latin scripts (Arabic, Chinese, etc.) require embedded fonts which pdf-lib doesn't have built in, so those may fall back or be skipped.
• Original PDF metadata — title, author, creation date are replaced with generic values on export.
• Password-protected scans — encrypted PDFs need to be unlocked before OCR.

Why browser-based OCR matters for privacy 🔒

Every other "free OCR online" service uploads your document. For a public poster, that's fine. For a scanned tax return, a medical report, a signed NDA, or a passport, uploading it to a stranger's server is a non-starter. You have no idea how long they keep the file, who has access, or what they do with the recognised text.

This tool works differently. The OCR engine is compiled to WebAssembly and downloaded to your browser — not the other way around. Your file never leaves the tab. Neither does the rendered image, the recognized text, or the searchable PDF you build. Everything is in memory on your device.

• No server uploads — your PDF never leaves your device.
• No server-side OCR — the engine runs locally on your CPU.
• No sign-up — no email, no account, no waiting.
• No page limits — the only constraint is your browser's memory.
• Works offline — after the first run, both engine and language data are cached.
• Free forever — no ads inside the tool, no per-file fee.

Who uses PDF OCR? 👥

• Researchers — making old papers and archives searchable for literature review.
• Lawyers — building searchable briefs from scanned exhibits and discovery documents.
• Accountants — indexing receipts and invoices for quick lookup later.
• Students — extracting quotes from scanned textbooks or historical sources.
• Journalists — pulling text out of scanned FOIA responses and leaked documents.
• Archivists and historians — turning image-only scans into searchable collections.
• Anyone with a scanner — making the paperless office actually searchable.

Common problems and how to fix them 🛠️

"My first run takes ages"

The first time you click Run OCR, the browser downloads the Tesseract WebAssembly engine (~4 MB) plus the language model. English is ~2.8 MB; Chinese and Japanese are larger. On a slow connection this can take a minute. After that, everything is cached and runs instantly.

"The extracted text is gibberish"

Three usual causes: (1) wrong language selected — if the PDF is in Spanish, using "English" produces nonsense; (2) too-low DPI — bump the rendering scale up; (3) the page is rotated 90° in the scan — Tesseract struggles with sideways text, so rotate the pages first with our Organize PDF tool.

"Tesseract only returns fragments"

This happens with multipage documents that have poor contrast or heavy backgrounds. Try turning on both Grayscale and Contrast boost. If that fails, re-scan the source at a higher DPI.

"The searchable PDF won't open in some viewers"

Chrome, Adobe Reader, macOS Preview, and Firefox all open pdf-lib output reliably. Very old viewers may not recognise the newer text-object features. Try opening the file in a current browser first.

"The invisible text doesn't line up with the visible text"

This can happen if the scan is skewed (a few degrees off horizontal). Tesseract reports word boxes in the skewed coordinate system, so the invisible text sits slightly off. The fix is at the source: straighten the scan before OCR.

"Non-Latin scripts (Arabic, Chinese) don't appear in the searchable PDF"

pdf-lib's standard fonts only cover Latin. For Arabic, Urdu, Hindi, Chinese, Japanese, and Korean, the recognized text will appear in the extracted .txt but not always in the invisible layer. The text is still in the PDF's internal data — some viewers can still search it — but rendering isn't guaranteed. For those languages, download the .txt file and pair it with the original scan.

PDF OCR vs PDF to Text vs PDF to Word 🤔

Tips for the best results 💡

• Straighten the scan first. A 3° skew is enough to hurt recognition. Use your scanner's auto-straighten or the rotate tool.
• Use 200–300 DPI. Higher than 300 rarely helps and slows everything down; lower than 150 makes OCR guess.
• Pick the right language. This is the single biggest accuracy lever after DPI.
• Test one page first. For a long document, set the page range to just page 1, run it, and check the text quality before committing to the whole file.
• Chunk very large PDFs. On mobile devices, Tesseract's memory footprint becomes a problem after ~60–80 pages. Process 25 pages at a time.
• Grayscale helps tinted paper. Old yellowed scans recognize significantly better after converting to grayscale.
• Keep the original scan. Don't overwrite it — the searchable PDF is a new file. If OCR made mistakes, you may want to try again with different settings.
• Mobile users: OCR works on phones but is slower. Desktop is recommended for documents over 20 pages.

Is it safe to OCR PDFs here? 🔐

Yes. Here's exactly what happens when you use this tool:

1. You pick a PDF from your computer.
2. Your browser reads it into memory locally.
3. PDF.js renders each page to a canvas — still in memory, still on your device.
4. Tesseract.js (WebAssembly) processes the canvas on your CPU — no network call.
5. The recognized text and searchable PDF are built in the browser and downloaded directly.

At no point does the file, the rendered image, or the recognized text leave your device. No data is sent to SkillsInsider's servers. No tracking cookies watch your OCR. You can prove it yourself: run one OCR pass to populate the cache, then turn off Wi-Fi and run another — it still works perfectly.