PDF OCR

Optical Character Recognition is the process of looking at an image of text and producing the actual text characters. It's a 60-year-old technology that's gotten dramatically better in the last decade thanks to neural-network approaches. Modern OCR engines achieve 95–99% accuracy on clean printed text, which makes them practical for everyday use: turning a phone photo of a recipe into editable text, extracting quotes from a scanned book, digitizing a stack of receipts.

This tool runs **Tesseract.js**, the JavaScript port of Google's Tesseract OCR engine (the open-source standard since version 4 in 2018). It works entirely in your browser via WebAssembly — no server, no upload, no cloud OCR service. The PDF is rendered page-by-page to a high-resolution image in your tab, then each image is fed to Tesseract, which returns the recognized text.

**The classic use case**: a scanned PDF where the text isn't really text. A 1970s contract that was photocopied so many times the original is lost. A receipt photographed with your phone. A whiteboard snapshot from a meeting. A book chapter you scanned for a class. In every case, the PDF visually shows text but the text isn't selectable, copyable, or searchable. OCR converts the pixels back into characters so the document becomes usable.

**Why a browser-based OCR tool?** Most online OCR services upload your PDF to their servers, run OCR there, send back the text. For most users that's fine; for sensitive documents (private correspondence, medical records, legal documents, financial statements), it's a privacy hole. This tool runs the entire pipeline locally — your PDF and its contents stay in your tab.

**The performance picture.** Tesseract.js needs to load two things on first use:

- **The Tesseract wasm binary** (~2 MB) — the OCR engine itself. - **A language data file** (~10 MB per language) — trained models for character recognition in a specific script.

Both are loaded lazily on the first OCR run and cached by the browser for the rest of the session. So:

- **First run**: 30–60 seconds for downloads, then 5–15 seconds per page for OCR. - **Subsequent runs (same session)**: 1–2 seconds startup, 5–15 seconds per page. - **Subsequent runs (later sessions, after browser cache eviction)**: back to first-run cold-start times.

For a 10-page document on first use, expect 2–3 minutes total. For the same document later in the same session, expect 1–1.5 minutes.

**Accuracy depends heavily on input quality.** The best inputs are:

- **High DPI**: 300 dpi minimum, 600 dpi ideal. Phone photos are typically 1500–3000 px wide which translates to roughly 200–400 dpi for letter-size documents — usable but not ideal. - **Flat scan**: paper-against-glass scans beat handheld photos because there's no perspective distortion and consistent lighting. - **High contrast**: dark text on white background. Color text or text on textured background reduces accuracy. - **Standard fonts**: Times, Arial, Helvetica, Calibri, Cambria, Georgia — anything Tesseract has seen a lot of. Decorative fonts, handwriting fonts, stylized lettering all degrade accuracy. - **Single-column layout**: like the PDF-to-Text tool, OCR results can jump between columns or include sidebar text inline.

For documents that don't meet these criteria, the output is still usable — you just have more cleanup work to do.

**Language support.** Ten languages exposed in the UI: English, Spanish, French, German, Italian, Portuguese, Russian, Chinese Simplified, Japanese, Korean. Tesseract itself supports 100+ languages (Arabic, Hebrew, Thai, Vietnamese, Hindi, and many regional scripts) — these aren't in the UI but could be added if there's demand. For mixed-language documents, pick the dominant language; Tesseract handles occasional foreign words reasonably well as long as they're in a script similar to the selected language.

**Progress and partial results.** OCR is processed page-by-page sequentially. The tool shows a progress bar (page 3 of 12) and updates the result textarea as each page completes. You can read partial output while later pages are still being processed. If you have what you need from the first few pages, the rest are still useful but you don't need to wait.

**Output format.** Plain text in the textarea, with optional `--- Page N ---` separators between pages. Copy to clipboard or download as a `.txt` file. The output is UTF-8 — works for Cyrillic, CJK characters, accented Latin, anything Tesseract recognizes.

**Privacy.** Tesseract.js runs entirely in your browser as WebAssembly. The PDF is rendered locally via pdf.js, the image data is fed to Tesseract locally, the text is generated locally. The only network requests are the initial CDN fetches of the Tesseract wasm and language data — these are cached after first load. Subsequent OCR jobs produce zero network traffic. Verify in DevTools.

**Edge cases handled:**

- **Network failure during initial download**: a clear error message indicates the CDN fetch failed; user can retry. Already-cached language packs don't re-download. - **Pages that are partially scanned + partially digital text**: OCR runs on the rendered page image and recognizes whatever pixels are there — including digital text on a partially-scanned page, which works fine because rendered digital text is just an image too. - **Pages with mostly images and a small text caption**: OCR will pick up the caption text and ignore the image regions. - **Rotated pages**: rendered with rotation applied, so OCR sees the upright orientation. (If your scan is upside-down, fix it first with PDF Rotate.) - **Multi-column scans**: column-jumping is common in OCR output; this is a Tesseract limitation, not a tool bug.

**What it doesn't do**: form-field recognition, table-structure recognition, math-formula recognition, handwriting (Tesseract is poor at handwriting), confidence scoring per character. For all of those, more specialized tools exist; Tesseract is the practical right answer for general-purpose printed-text OCR.