Skip to content

v2.1.0

Latest

Choose a tag to compare

@github-actions github-actions released this 24 May 16:08

v2.1.0

New Features

  • OCR Language Option: Added Chinese/mixed Chinese-English and English language modes for PaddleOCR recognition.
    • GUI: Language dropdown in both single-file and batch task dialogs.
    • CLI: New --ocr-lang argument (ch or en, default ch).
    • English now explicitly uses en_PP-OCRv5_mobile_rec, avoiding the issue where lang was silently overridden by explicit model names.
  • OCR Preprocessing: All OCR input is now converted to grayscale and uniformly scaled (longest side → 1920 by default) before recognition, improving text-box detection stability for PPT/screen-sized documents.
    • New environment variable MINERU_OCR_MAX_SIDE controls the scaling threshold; backward-compatible with the old MINERU_OCR_CPU_MAX_SIDE.
    • Coordinates are scaled back to original dimensions after OCR; rendering and refinement still use the original RGB image.
  • Release Manifest: Each release now includes a release-manifest.json with version, commit, SHA-256 checksums and sizes for all installer assets. Hyphenated version tags (e.g. v2.1.0-rc.1) are automatically marked as pre-releases.

Improvements

  • OCR Advanced Labels: Threshold labels in the GUI now show default values and recommended ranges (e.g. "Det Thresh, default 0.35, range 0.1–0.9").
  • Watermark Detection: discarded_blocks from MinerU are no longer all treated as watermarks. Only blocks containing NotebookLM text are marked as watermarks, preventing accidental removal of headers and footers.
  • MinerU Block Handling: Added chart to image-like block types. Unknown block types now default to image-like handling, preserving their visual footprint rather than crashing on empty text.

Bug Fixes

  • Fixed Text IR element requires text or text_runs crash caused by chart and unknown MinerU block types being treated as text.
  • Fixed over-aggressive watermark removal that deleted headers/footers from discarded_blocks.
  • Fixed OCR language option (--ocr-lang) not actually affecting PaddleOCR recognition because explicit model names overrode the lang parameter.

Testing

  • Unit tests: 25 passed
  • Integration tests: 27 passed
  • All demo case conversions verified

Full Changelog: v2.0.0...v2.1.0