Skip to content

v0.2.0 — PP-OCRv6 ONNX Unified Engine

Choose a tag to compare

View all tags
@github-actions github-actions released this 25 Jun 14:52
· 120 commits to main since this release
v0.2.0
7d38c5f

v0.2.0 — PP-OCRv6 ONNX Unified Engine

Pure ONNX recognition engine — no Zinnia, no Tegaki, no proprietary engines.

What's New

  • PP-OCRv6 ONNX unified engine replacing all legacy backends with a single deep-learning engine
  • 18,710 character CJK recognition via PP-OCRv6 MobileNetV3 small model
  • 100% top-1 accuracy validated on 40 real handwriting samples
  • MAX-pooled confidence scoring fixing the confidence dilution bug in CTC decoder

Package Improvements

  • Cross-distro support: .deb (Debian/Ubuntu), .rpm (Fedora/openSUSE), source tarball (Arch)
  • ONNX model auto-downloaded on install — not stored in the repository
  • pip-based onnxruntime for all distros

Bug Fixes

  • Fixed dict index corruption from line.strip() stripping U+3000 ideographic space
  • Fixed confidence pooling to use MAX instead of mean across CTC time steps
  • Fixed stroke line width from 6 to 8 for training-data alignment
  • CI: ONNX model download, pip deps, --break-system-packages fallback, zypper --no-refresh, archlinux python-cairo

Assets

  • ibus-handwrite-chinese_0.2.0_all.deb — Debian/Ubuntu package
  • ibus-handwrite-chinese-0.2.0-1.fc44.noarch.rpm — Fedora/openSUSE package
  • ibus-handwrite-chinese-0.2.0.tar.gz — Source tarball
Assets 5
Loading