Offline OCR

Offline OCR — A safe, fully offline region-select OCR Select any area on screen with your mouse to recognize text and copy it to your clipboard instantly. Powered by NDLOCR, an OCR engine developed by the National Diet Library of Japan (NDL), delivering high-accuracy Japanese text recognition. All processing happens entirely in your browser. No internet connection required — your images and text are never sent to any external server. ■ Key Features - Fully offline — your data never leaves your device - One-click usage — just select a region and text is automatically copied - Optimized for Japanese — powered by the National Diet Library's OCR technology - 3 ways to launch — toolbar icon / keyboard shortcut / right-click menu - Regex cleaning rules — auto-remove commas, spaces, or line breaks before copying. Create your own custom rules to fit your workflow, such as stripping digit separators from invoice amounts - Image viewer — open local images or clipboard images for OCR, with zoom and region selection - PDF viewer — open local PDFs and OCR any page (v0.8.0+) - Deskew & select — for angled photos or documents with visible perspective, fit a quadrilateral with four draggable corners and OCR the unwarped result, with a live "corrected preview" updating as you drag (v0.9.0+) - Multilingual UI — automatically switches between Japanese and English ■ How to Use 1. Click the extension icon (or use the keyboard shortcut) 2. Drag to select the area you want to OCR 3. Wait a few seconds — the result is automatically copied to your clipboard 💡 Image viewer You can also OCR local image files and clipboard images. - From the extension popup: "Open Image" / "Open Clipboard Image" - Right-click an image on any web page → "Open Image in Viewer" In the viewer you can zoom in to select a region, or OCR the entire image at once. ■ Great For - Copying text from PDFs or images - Quickly extracting numbers from invoices and receipts - Transcribing text from screenshots - Working with confidential documents you don't want to upload to external services ■ Privacy Offline OCR makes absolutely no network requests. The OCR engine (ONNX Runtime Web) and models are fully bundled within the extension and run entirely inside your browser. No personal data or images are ever collected or transmitted. ■ Technical Details - OCR Engine: NDLOCR-Lite (lightweight version of the National Diet Library OCR) - Detection Model: DEIM (FP32) - Recognition Model: PARSeq (FP32) - Inference Runtime: ONNX Runtime Web (WebAssembly) - Supported Languages: Japanese (horizontal text) - Requirements: Chrome 116 or later ■ Acknowledgments The OCR engine and models used in this extension are based on NDLOCR, researched, developed, and published by the National Diet Library of Japan (NDL). We sincerely thank the National Diet Library for making their high-accuracy Japanese OCR technology openly available. ■ Changelog v0.9.0 - Added "Deskew & Select" mode. For angled photos or documents with visible perspective, drag the four corner handles in the viewer to fit the region — it's then unwarped with a perspective transform before OCR - Live "corrected preview" in the top-right updates as you drag, so you see exactly what will be fed to OCR before committing - Press Enter to confirm (works alongside the on-screen button) - Corner drags are constrained to keep the quad convex, so the selection is always a valid region - Output resolution is chosen automatically from the four edge lengths so small text isn't compressed - The unwarped image flows through the same downstream pipeline as rectangular selection, so all prior accuracy work (v0.7.1 row-ink profile, v0.8.1 CTC loop detection, dark-bg inversion) applies with no extra plumbing - Zero-network policy preserved (perspective math is pure TypeScript, no new external libraries) v0.8.1 - Faster OCR on multi-line pages. PARSeq recognition now runs up to 4 lines in parallel via a worker-pool (newspaper columns, receipts, multi-paragraph documents) - Output order is unchanged. No effect on 1–2 line captures - Fixed a bug where OCR'ing a short Latin label on a dark background could return "the the the ..." (a repeated token) instead of the actual text v0.8.0 - PDF OCR support. From the popup's new "Open PDF" button, open a local PDF in the viewer and OCR any page you like (password-protected PDFs are not yet supported) - The viewer now has page navigation (◀ / ▶ buttons, page indicator, ← / → keys) - OCR history entries are labeled `filename.pdf#p2` so you can trace which page of which PDF a result came from - All PDF rendering and OCR runs locally — the zero-network policy is preserved v0.7.1 - Fixed a bug where OCR'ing a small selection could append unwanted trailing characters (a run of "1"s) to the result - Fragments of adjacent lines were the culprit; padding color is now taken from the majority color across the border, and a row-ink profile step picks the main text band while dropping thin fragment bands v0.7.0 - Added OCR history. Every clipboard copy is automatically saved locally on your device (data is never sent externally); re-copy, edit, delete, and search from the new history page - Open the history via "Open history" in the popup or on the options page v0.6.2 - Properly fixed the v0.6.0 issue that v0.6.1 had rolled back, and re-enabled the furigana filter and small-region OCR features v0.6.1 - Emergency rollback to v0.5.0 behavior. The v0.6.0 release had a bug where OCR'ing short alphanumeric text could append a long run of unwanted characters (1s, 9s) to the result v0.6.0 - Automatically filter out furigana (ruby text) in manga and annotated text — only the main body text is returned - Small selections like button or link text are now recognized v0.5.0 - OCR output for tables and receipts is now tab-separated (paste into Excel / Google Sheets and each cell lands in its own column) - Cleaning rule replacement strings now support escape sequences (\t …