paddleocr-onnx-models

PaddleOCR ONNX Models

Pre-converted PaddleOCR models in ONNX format for use with Kreuzberg.

Models

Pipeline

PaddleOCR uses a three-stage pipeline:

1. Detection (det): Locates text regions in the image using a DB (Differentiable Binarization) network. The detection model is language-agnostic — it identifies bounding boxes around text regardless of script.
2. Classification (cls): Determines text orientation (upright vs upside-down) to correct rotated text before recognition.
3. Recognition (rec): Reads characters from each detected text region. This is the only language-specific model.

Sources & Credits

These models are derived from PaddlePaddle/PaddleOCR, licensed under Apache 2.0.

• Detection model (ch_PP-OCRv4_det_infer.onnx): Pre-converted ONNX model sourced from SWHL/RapidOCR on HuggingFace (PP-OCRv4 folder). Originally trained by PaddlePaddle as ch_PP-OCRv4_det_infer.
• Classification model (ch_ppocr_mobile_v2.0_cls_infer.onnx): Pre-converted ONNX model sourced from SWHL/RapidOCR on HuggingFace (PP-OCRv1 folder). Originally trained by PaddlePaddle as ch_ppocr_mobile_v2.0_cls_infer.
• Recognition model (en_PP-OCRv4_rec_infer.onnx): Converted from PaddlePaddle format to ONNX using paddle2onnx (opset 14). Original PaddlePaddle model downloaded from PaddlePaddle's official distribution (en_PP-OCRv4_rec_infer).

Acknowledgments

• PaddlePaddle for creating and training the PP-OCR model series
• SWHL/RapidOCR for maintaining pre-converted ONNX models
• Paddle2ONNX for the PaddlePaddle-to-ONNX conversion tool

License

The original PaddleOCR models are licensed under Apache License 2.0.