OnnxOCR

Multilingual OCR that runs on ONNX Runtime alone

The upstream project is jingsongliujing/OnnxOCR.

This package is built from KumaTea/OnnxOCR.

Release lines

Each major version is one model generation:

To stay on one generation, pin the major version, e.g. pip install "onnxocr<4".

Install

pip install onnxocr

Requires Python 3.11+. The wheel (~75 MB) bundles the PP-OCRv6 small and tiny detection and recognition models, the PP-OCRv5 models and the angle classifier.

Optional extras:

Usage

import cv2
from onnxocr.onnx_paddleocr import ONNXPaddleOcr

model = ONNXPaddleOcr(use_angle_cls=True, use_gpu=False)

img = cv2.imread("test.jpg")
result = model.ocr(img)

for box, (text, score) in result[0]:
    print(f"{score:.3f}  {text}")

Save an annotated image:

from onnxocr.onnx_paddleocr import sav2Img

sav2Img(img, result, name="result.jpg")

Non-ASCII paths on Windows need cv2.imdecode rather than cv2.imread:

import numpy as np
img = cv2.imdecode(np.fromfile(path, dtype=np.uint8), cv2.IMREAD_COLOR)

Model size

PP-OCRv6 comes in three sizes. This package defaults to small; upstream defaults to medium.

model = ONNXPaddleOcr(ocr_model_size="tiny")

To use medium, download and extract PP-OCRv6_medium_det and PP-OCRv6_medium_rec from PaddleX, then pass their paths:

model = ONNXPaddleOcr(
    ocr_model_size="medium",
    det_model_dir="PP-OCRv6_medium_det_onnx_infer/inference.onnx",
    rec_model_dir="PP-OCRv6_medium_rec_onnx_infer/inference.onnx",
)

Other modes

ONNXPaddleOcr also exposes license-plate, table and layout recognition via use_plate_recognition=True, use_table_recognition=True and use_layout_analysis=True. Those models are not bundled — download them first:

python scripts/download_models.py --source huggingface

See the upstream README for those modes, the HTTP API service, the WebUI and Docker images.