PP-OCRv6 medium text detection (ppocr-det)
Text line detection for manuscript pages, served by the nomikos
ppocr-det runtime adapter on ONNX Runtime CPU. This repo holds the
converted detection graph plus this card; the adapter, registry pin and
validation live in the nomikos repository.
| Field | Value |
|---|---|
| registry model id | ppocr-segment |
| registry tag | stable |
| task | segment |
| weights source | hf://nomikos-project/segmentation-ppocrv6-det@stable |
Source and licence
Converted from PaddlePaddle/PP-OCRv6_medium_det at Hub revision
8e0f56fb2ef86b461d99cfc7ac5c137738985f61, whose licence is Apache-2.0.
Redistribution of converted weights under that licence is permitted, and
this repo is published under the same terms.
Conversion
Converted with paddle2onnx at opset 21 (ai.onnx) from the PaddleX model
directory, with dynamic batch, height and width (no axis frozen).
onnx.checker.check_model passes. The served file is ppocrv6-det.onnx:
| File | SHA-256 | Size (bytes) |
|---|---|---|
ppocrv6-det.onnx |
09e4c827c5bb20a0344374bbf8b88d41b7c8bf2be3a0db82ffed3bf090eacfe3 |
62064678 |
Input x is [N, 3, H, W]; output is [N, 1, H, W].
Serving parameters
Preprocessing reproduces PaddleX 3.7.0: PIL decode to RGB, BGR channel
order, scale down only when the longer side exceeds limit_side_len
(default 1920, each side rounded to a multiple of 32) with bilinear
resize, ImageNet normalisation, CHW float32 batch tensor. DB postprocess
mirrors PaddleX 3.7.0 in quad mode: thresh 0.2, box_thresh 0.45,
max_candidates 3000, unclip_ratio 1.4 with pyclipper JT_ROUND,
use_dilation false, score_mode fast. Recommended session options are
ORT_ENABLE_EXTENDED, sequential execution, 1 inter-op thread and 4
intra-op threads (override with NOMIKOS_PPOCR_DET_THREADS).
Parity result
The ONNX graph reproduces Paddle static inference exactly: 28 of 28
page and cap cases with 0 boxes gained or lost and 0.000 px corner
distance, and the serving adapter reproduces the Paddle pipeline boxes at
0.000 px on all 14 reference pages. Full evidence is in the nomikos
repository under docs/inference/ppocrv6-onnx-parity-2026-09-19.md and
docs/inference/ppocrv6-segmenter.md.
Benchmark
Coptic line localisation, 12 pages and 1018 target lines: P 90.6, R 99.0, F1 94.6, against BLLA at F1 71.1.
Known gaps
No region or column blocks (every line hangs under one full-page block), enlarged initials can be separate boxes, neighbouring boxes can overlap, watermarks and page labels and marginal text are returned as ordinary lines, and the model is not the default segmenter until complete OCR crops are validated.