handwritten-ocr

Photos of handwritten Chinese notes and math formulas are often poorly handled by general-purpose OCR services, and sending them to a third-party cloud is uncomfortable when privacy matters. The handwritten-ocr plugin brings the entire recognition pipeline onto your local machine: handwritten Chinese text is transcribed into plain text, math formulas are converted to LaTeX, and the result is delivered as a clean Markdown document. Everything runs offline; no image or data ever leaves your computer.

Under the hood, several open-source models work together: Pix2Text MFD 1.5 detects formula regions, RapidOCR reads Chinese text lines, and UniMERNet (with Pix2Text MFR 1.5 as a fallback) converts formulas to LaTeX. Three inference backends are available-GPU via DirectML, CPU, and an experimental NPU mode-while all tunable parameters live in a single settings.json file with hot-reload support, so no code changes are needed. You can drive the plugin through a FastAPI service with a browser-based admin panel, a CLI script for batch processing, or the ocr_image tool preset in DSH, letting an LLM call the recognizer mid-conversation and receive Markdown back.

If you regularly take notes by hand in lectures, meetings, or research sessions and want to turn that paper into searchable, editable text, this plugin is built for you. It is especially well suited to researchers, students, and developers who need private, on-device OCR or who want to integrate handwritten-image recognition seamlessly into LLM tool frameworks like DSH. One installation, fully offline, ready to use.