ppu-paddle-ocr is stable and in active maintenance. This roadmap states the direction and the near-term priorities. It is a statement of intent, not a contract; dates are deliberately omitted because a small team sets pace by capacity.
- Keep dependencies current and the supply chain hardened: Dependabot updates, pinned actions, npm provenance, the SCA and SAST gates in CI.
- Fix reported bugs and answer issues within the timeframes in SECURITY.md.
- Hold test coverage steady (enforced in CI).
- Broaden language and script coverage as PP-OCR models and dictionaries allow (Latin, Cyrillic, Arabic, Indic, CJK, Thai), driven by real demand.
- Sharpen the recognition strategies (
per-box/per-line/cross-line) and the engine choice (opencv/canvas-native) with benchmarks. - Improve the WebGPU path and its CPU fallback on the browser entry point.
- Performance work on detection and recognition hot paths, with before/after
benchmarks (
bun task bench). - Evaluate newer ONNX Runtime releases as upstream ships them.
- Grow the
apps/serveHTTP service and CLI surface where it helps real users.
- Training or fine-tuning models. This project runs inference on published PP-OCR models; training lives elsewhere.
- Features outside OCR (general computer vision lives in sibling packages such as ppu-ocv).
Open an issue to discuss direction, or a pull request for a concrete change. See CONTRIBUTING.md and GOVERNANCE.md.