Document OCR & Searchable PDF Studio

OCR (Optical Character Recognition) analyzes pixel patterns in images to recognize glyph shapes and translate them into machine-readable digital characters. This tool compiles the open-source Tesseract OCR engine to WebAssembly, running entirely inside a browser Web Worker without external cloud dependencies.

The tool supports raster images in JPEG (.jpg, .jpeg), PNG (.png), and WebP (.webp) formats, as well as multi-page PDF documents. All rendering and extraction take place locally in browser memory.

Rather than passing raw PDF binaries directly to an OCR engine, the tool utilizes PDF.js to render each PDF page into a high-resolution Canvas bitmap, executes OCR on that rendered page image, and then reconstructs a dual-layer Searchable PDF using pdf-lib.

A digital text PDF contains native font vectors and embedded character encodings created by word processors. A scanned PDF is merely a collection of container pages holding static raster photographs where text cannot be selected or searched until OCR is applied.

A Searchable PDF combines the exact original scan image on the visual layer with an invisible, selectable Unicode text layer situated directly above the corresponding words. Users see the original document while enjoying native search and copy capabilities.

The OCR engine computes bounding boxes (x0, y0, x1, y1) for every recognized word and line. These pixel coordinates are mathematically mapped to the PDF point coordinate system (72 DPI) with vertical axis inversion so the invisible text aligns precisely over the printed letters.

OCR confidence is an engine-reported statistical metric (0 to 100) reflecting how closely a recognized glyph matches trained language dictionaries. No OCR system can guarantee 100% accuracy due to potential blur, skew, or non-standard typefaces; the tool displays raw, objective confidence values without false claims.

Words with confidence scores below 60 are flagged as 'Possible OCR uncertainty'. They are highlighted on the review canvas so users can inspect and correct ambiguous words rather than blindly trusting the machine output.

Yes. The studio features an interactive review panel where you can edit line text or individual word tokens directly. All manual edits immediately propagate to the generated Searchable PDF and all exported data files.

Physical document scans frequently suffer from skew, shadow gradients, low contrast, and scanner noise. Preprocessing operations like automatic deskew, Otsu binarization, and sharpening clean up character edges, significantly reducing misrecognition rates.

A setting of 150 to 200 DPI offers an optimal balance between recognition accuracy and memory consumption for standard documents. For fine print or complex scripts, 300 DPI can be selected, though it requires more memory.

The local OCR model is primarily optimized for horizontal text flows. Vertical text recognition in Japanese or Chinese is limited and may result in split lines or non-standard reading orders.

The engine is trained specifically on printed typographic fonts. Cursive or informal handwriting is not specifically optimized and typically yields low confidence; manual transcription review is strongly recommended.

The tool detects multi-column text flows and marks tabular regions as 'Table candidates'. However, it does not promise complete spreadsheet cell reconstruction or flawless layout reproduction for multi-column magazines.

Alphanumeric serial numbers and financial codes often confuse visually identical glyphs like '0 / O', '1 / I / l', '5 / S', and '8 / B'. The engine automatically flags these mixed sequences as potential ambiguities for human verification.

Yes. Detected or selected text tokens can be replaced with '[REDACTED]'. In addition, applying image redaction permanently burns solid black rectangles into the exported PDF image, ensuring underlying pixels cannot be retrieved.

Never. The application enforces a strict zero-network, zero-upload invariant. No files leave your device, and no persistent browser storage (IndexedDB, localStorage, cookies) is utilized.

Yes. All WebAssembly binaries, worker scripts, and trained language models are bundled directly with the application. Once the web application is loaded, it executes flawlessly with network access completely disabled.

Yes. The UI language (e.g., English, Japanese) is completely decoupled from the OCR language. You can freely select between English, Japanese, Simplified Chinese, German, French, Spanish, Arabic, Russian, Hindi, and Portuguese.

Immediately after building the PDF, the tool uses an isolated instance of pdfjs-dist to reopen the binary, extracting text content and verifying that page counts match and invisible text layers are discoverable.

Yes. The interface is fully responsive across mobile (375px), tablet (768px), and desktop (1440px) viewports with stacked preview panes. However, for multi-page high-DPI document processing, a desktop browser is recommended.

To prevent browser memory exhaustion, processing is performed page-by-page with immediate canvas buffer disposal. The maximum supported file size is 50MB with a limit of 100 pages per document.