az-ocr

自动收集自 crates/algorithm/ocr/README.md。

OCR backends for the addzero ecosystem.

The crate currently exposes two backend families:

document: PP-StructureV3 document parsing bridge for complex layout, tables, and Markdown output.
paddle: the existing local PaddleOCR ONNX detection and recognition path.
tesseract: a wrapper around the system tesseract command through rusty-tesseract.

rusty-tesseract does not ship OCR model weights. It delegates to the installed Tesseract binary and whatever language .traineddata files are available in the active tessdata directory.

The document backend delegates complex table, formula, reading-order, layout, and Markdown conversion to PaddleOCR's official PP-StructureV3 Python API. It standardizes the outputs as:

document.md
structured.json
manifest.json

For production Chinese OCR quality, keep PaddleOCR as the main path. Tesseract is available as a secondary backend for environments where its local traineddata packages are acceptable. The intended upgrade path is:

use PP-OCRv6 or newer PaddleOCR models for text detection and recognition,
use PP-StructureV3 for complex table/layout/formula/reading-order recovery,
normalize document parsing outputs through the document backend instead of trying to infer tables from plain OCR text lines.