Local Czech Book OCR

Local OCR pipeline for sequential JPEG scans of Czech literary book pages.

Setup

Install a CUDA-enabled PaddlePaddle build appropriate for the machine, then install the project dependencies. PaddleOCR currently uses PaddlePaddle rather than PyTorch for the traditional OCR stage.

Example:

uv sync

If the selected PaddlePaddle wheel is not available through the default index for the installed CUDA version, install PaddlePaddle manually using the official PaddlePaddle instructions, then install this project.

Usage

Dry-run page discovery:

uv run book-ocr --input-dir /dev/shm/milena --dry-run

Run OCR:

uv run book-ocr --input-dir /dev/shm/milena --output-dir output

Main output:

• output/raw_manuscript.txt: one continuous raw OCR text body.
• output/pages/page_000001.txt: raw text for each page.
• output/pages/page_000001.json: OCR records for each page.
• output/pages.jsonl: all page OCR records.
• output/manifest.json: input file manifest.
• output/report.json: basic quality statistics.

Notes

• Czech language is configured with PaddleOCR language code cs.
• The first pass intentionally preserves raw line breaks and punctuation. Cleanup, dehyphenation, and paragraph merging belong in the later VLM stage.
• More JPEG files can be added later; rerun the same command to rebuild outputs from the current directory contents.