Skip to content

PP-OCRv5

PaddlePaddle's pretrained PP-OCRv5 text recognition and detection, tried as an alternative to SimpleHTR and TrOCR (issue #157). The mobile English recognition model and the mobile detection model are used as they are, run as ONNX through rapidocr.

Exception to ADR 006: the weights are not our own export; rapidocr downloads them on first use. The dependency sits behind the optional ppocr extra.

Two pipelines, to answer two questions:

Pipeline Question
2026-10-09_yolo_detector_ppocrv5 Is PP-OCRv5 a better recogniser than SimpleHTR? YOLO detector unchanged, so precision and recall are identical to the other YOLO pipelines.
2026-10-09_ppocrv5_det_rec Does PP-OCRv5 work as a full pipeline? Its own line detector, split into words by rapidocr.

GPU training setup

Not applicable: we have not trained or finetuned PP-OCRv5.

Inference

uv sync --extra ppocr

from xournalpp_htr.inference_models import PPOCRv5Model

model = PPOCRv5Model.from_pretrained()
texts = model.recognize_batch(word_images_grayscale)  # recognition only
words = model.read_page(page_image_grayscale)  # [(text, BoundingBox), ...]

Benchmark:

uv run --extra ppocr python scripts/run_benchmark.py -p 2026-10-09_yolo_detector_ppocrv5
uv run --extra ppocr python scripts/run_benchmark.py -p 2026-10-09_ppocrv5_det_rec

Best model

Neither pipeline replaces the default. The recognizer-only pipeline 2026-10-09_yolo_detector_ppocrv5 has the lowest CER of all pipelines (28.4% case-insensitive) but a lower word accuracy (42.0%) than SimpleHTR (46.7%) and TrOCR large (48.5%); see Experiments.

Experiments

2026-10-09 — PP-OCRv5 as recogniser and as full pipeline

  • Hypothesis: PP-OCRv5 beats SimpleHTR (CER 31.7%, word accuracy 46.7%, 2026-10-05_yolo_detector_beam_vocab) on handwritten word crops, and possibly works without our YOLO detector.
  • Setup: benchmark dataset latest; code revision 712027d; CPU only. Mobile English recognition model and mobile detection model of PP-OCRv5.
  • Command: uv run --extra ppocr python scripts/run_benchmark.py -p <pipeline>.
  • Results:
Pipeline Precision Recall CER (case-insens.) CER (case-sens.) R×(1-CER) Word Acc
2026-10-05_yolo_detector_beam_vocab (SimpleHTR, for reference) 73.8% 80.1% 31.7% n/a n/a 46.7%
2026-10-09_yolo_detector_trocr_large_strip_punct (for reference) 73.8% 80.1% 29.3% 37.9% 56.6% 48.5% (82/169)
2026-10-09_yolo_detector_ppocrv5 (recognizer only) 73.8% 80.1% 28.4% 29.4% 57.3% 42.0% (71/169)
2026-10-09_ppocrv5_det_rec (detector + recognizer) 71.1% 50.2% 29.8% 30.2% 35.2% 42.5% (45/106)

The last row's CER and word accuracy cover only the 106 matched words, so they flatter it. - Best model: 2026-10-09_yolo_detector_ppocrv5 — CER 28.4% (case-insensitive), R×(1-CER) 57.3%, but not the overall best: its word accuracy is below SimpleHTR and TrOCR large. Not the pipeline of record. - Conclusion: - As a recogniser, PP-OCRv5 (8 MB, ONNX) has the lowest CER, but the gain over SimpleHTR (about 3 points) is likely within noise on 169 words, and word accuracy is lower. No reason to replace SimpleHTR. - As a full pipeline it finds far fewer words than our YOLO detector (recall 50.2% vs. 80.1%), so detection is the weak part. - Next: the server variants and PP-OCRv6, if worth the effort.

Current status

  • PPOCRv5Model implemented with recognize, recognize_batch and read_page; both pipelines are wired into compute_predictions.
  • Benchmarked; see Experiments.
  • PP-OCRv5 is trained mostly on printed text, so handwriting accuracy is uncertain. The online demo check (stage 1 of issue #157) was done separately.
  • 2026-10-09_ppocrv5_det_rec is not in CROP_RECORDING_PIPELINES: it has no per-word crops to record.

Outlook

  • Try the server variants and PP-OCRv6 (both supported by rapidocr).