完美排版ocr

Full OCR pipeline for scanned PDFs with layout preservation. Use this skill whenever the user wants to OCR a PDF, convert a scanned document to searchable text, or preserve the original layout of a scanned book/document. Triggers on: "OCR this PDF", "用PaddleOCR处理", "识别这个PDF", "扫描版PDF转文字", "把这个PDF做OCR", or when a PDF path is provided alongside any mention of OCR, text recognition, or layout preservation.

PDF OCR with Layout Preservation

Automated pipeline: Split → OCR API → Layout PDF → Merge

Each original page becomes one PDF page, with text placed at exact bounding-box positions and font sizes calibrated to fill the original block dimensions.

Quick Start

bash

python ~/.claude/skills/pdf-ocr-layout/scripts/pipeline.py "/path/to/input.pdf"

Output: input_ocr.pdf in the same directory. Intermediate files in input_ocr_work/.

Full Options

bash

python ~/.claude/skills/pdf-ocr-layout/scripts/pipeline.py \
  "/path/to/input.pdf" \
  --output "/path/to/output.pdf" \
  --work-dir "/path/to/workdir" \
  --chunk-size 90

Steps for Claude

1. Ask for the PDF path if not already provided in the conversation.
2. Check dependencies (install only what's missing):

   bash

   pip install pypdf reportlab Pillow requests -q
   

3. Run the pipeline and stream output to the user:

   bash

   python ~/.claude/skills/pdf-ocr-layout/scripts/pipeline.py "{input_pdf}"
   

4. Monitor progress — the script prints step-by-step progress including API polling. API jobs typically take 1–5 minutes per 90-page chunk.
5. Report the output path when done.

Resume / Retry

The pipeline saves state to the work directory and is fully resumable:

• jobs.json — API job IDs (prevents re-submitting already-queued chunks)
• chunk_*_results.jsonl — cached OCR results (skip re-downloading)
• chunk_*_ocr.pdf — completed chunk PDFs (skip re-rendering)

If interrupted, simply re-run the same command. It picks up where it left off.

Common Issues

Problem	Fix
ModuleNotFoundError	Run the pip install command above
API 4xx error	Check the PDF isn't password-protected
Job stuck in running	Normal for large chunks; wait up to 10 min
Missing images in output	Images left blank per design (API images are optional)
Font too small/large	The font size auto-calibrates — first page may look different if it's a cover

Output Quality

• Block positions: exact (scaled from 812×1269px OCR space to A4)
• Font sizes: auto-calibrated using fs = min(√(h×w / n×0.65), h×0.72) — verified to recover original ~13–14pt body text
• Page numbers, headers, footers: included (all block types preserved)
• Images: embedded if URL accessible, blank if not
• 1 OCR page = 1 PDF page: always maintained