Scanned PDF to Ebook Converter

Problem: Scanned PDFs Are Hard to Turn Into Editable Ebooks

Many older books, papers, and course notes exist as scanned PDFs where text lives inside images. Copying, searching, restructuring, and reading on e-book devices become awkward. Simply opening the PDF is not enough when the goal is a clean Markdown file or an EPUB package. Direct conversion often loses heading hierarchy, breaks tables, garbles formulas, or leaves images scattered. This skill targets that specific conversion path: from scanned page images to structured, editable book content, especially for Chinese and English technical or literary documents.

How It Works: Local OCR and Multi-Backend Conversion

The skill treats the PDF as a rendered document rather than a plain text stream. It first checks whether the PDF has a usable text layer. If pages contain enough characters, it extracts text directly and skips OCR. For scanned pages, it renders the page and runs a local OCR backend. Supported backends include pdf-craft, Marker, and MinerU, so you can choose based on Chinese quality, formula support, table fidelity, speed, or memory usage.

Key outputs include:
- Markdown: preserves heading structure, extracts images into an assets directory, and works well for re-editing or version control.
- EPUB: builds a table of contents, renders tables as HTML, and represents formulas as MathML, making it suitable for e-book readers.
- Batch conversion: processes a directory, skips existing outputs, and can resume after interruption.

The workflow relies on a local Python environment, PDF rendering tools, and OCR models. When a GPU is available, larger models can be selected; on CPU-only systems, it falls back to lighter settings. Apple Silicon machines can use MPS acceleration.

Boundaries

The skill is optimized mainly for Chinese and English, with reasonable support for Japanese and Korean. EPUB output is currently supported only by the pdf-craft backend. If formulas, tables, and code blocks are the priority, MinerU may be a better fit. If batch speed matters more, Marker is worth trying. Poor scans, heavy watermarks, skewed pages, handwritten notes, or complex cross-page tables can reduce accuracy. Treat the output as a draft: manually verify footnotes, headers, footers, formulas, and any publication-sensitive layout.