RapidOCR-ONNXRuntime

Warnings

• gotcha The `rapidocr-onnxruntime` package (PyPI version 1.4.4) is a core component and backend for the broader `RapidOCR` project on GitHub, which uses a `v3.x.x` versioning scheme. The user-facing API is primarily accessed through the `rapidocr` package (e.g., `from rapidocr import RapidOCR`), which integrates this ONNX Runtime backend.
  Fix: Always install `rapidocr-onnxruntime` if you intend to use the ONNX Runtime backend, and import the `RapidOCR` class from the `rapidocr` package for typical usage. Refer to the official RapidOCR GitHub documentation for the most current usage patterns.
• gotcha For GPU acceleration, installing `onnxruntime-gpu` is necessary. Simply installing `rapidocr-onnxruntime` will often default to `onnxruntime` (CPU). Additionally, `onnxruntime-directml` (which might be the default on some Windows systems) can lead to slower performance on older hardware.
  Fix: Explicitly install `onnxruntime-gpu` for GPU inference (e.g., `pip install rapidocr-onnxruntime onnxruntime-gpu`). For optimal performance, especially on Windows, consider the specific `onnxruntime` variant that best suits your hardware.
• gotcha While RapidOCR supports Chinese and English by default, using other languages or custom models often requires specific configurations or 'self-service conversion' of models. Directly using models not from the RapidAI/RapidOCR repository may lead to compatibility issues or require manual patching of configurations.
  Fix: Consult the RapidOCR documentation's 'Model List' and 'User Guide' sections for details on supported languages and the process for integrating custom or additional language models.
• gotcha Some users have reported that the `TextDetector` component within RapidOCR, when using the ONNX Runtime backend, can be slower for text detection tasks compared to PaddleOCR's native implementation.
  Fix: If text detection speed is critical and appears to be a bottleneck, consider experimenting with different `onnxruntime` configurations (e.g., thread counts) or evaluating alternative OCR engines, especially if migrating from PaddleOCR.

Install

• pip install rapidocr-onnxruntime Default installation

Imports

• RapidOCR

  from rapidocr import RapidOCR

  The primary user-facing class `RapidOCR`, which internally uses `rapidocr-onnxruntime` as its ONNX backend, is typically imported from the higher-level `rapidocr` package. Direct imports from `rapidocr_onnxruntime` are not for general API use.

Quickstart

This quickstart demonstrates how to initialize the RapidOCR engine and perform OCR on an image. The `RapidOCR` class handles model loading and inference using the ONNX Runtime backend provided by `rapidocr-onnxruntime`. The example downloads an image from a URL, processes it, and prints the detected text along with bounding box information.

from rapidocr import RapidOCR
from PIL import Image
import requests
from io import BytesIO

# Initialize the OCR engine
engine = RapidOCR()

# Example image URL (replace with your image path or URL)
img_url = "https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/resources/test_files/ch_en_num.jpg"

# Download and open the image
response = requests.get(img_url)
image = Image.open(BytesIO(response.content))

# Perform OCR
result = engine(image)

# Print the OCR results
print(result)

# Optionally, visualize the result (requires OpenCV installed and a display environment)
# import cv2
# import numpy as np
# image_np = np.array(image)
# for box, text, score in result:
#     box = np.array(box).astype(np.int32).reshape((-1, 1, 2))
#     cv2.polylines(image_np, [box], True, (0, 255, 0), 2)
#     cv2.putText(image_np, text, (box, box - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
# cv2.imshow("OCR Result", image_np)
# cv2.waitKey(0)
# cv2.destroyAllWindows()