How to use PaddleOCR from code using onnxruntime?

[thebest132]

I have downloaded the detection and recognition .onnx files. How do I use them from code? I would like to change the det_db_unclip_ratio parameter as well.

[GreatV]

You can use PaddleOCR with onnxruntime by loading the .onnx models for text detection and recognition. Below is a step-by-step guide on how to achieve this:

1. Install Required Dependencies

First, you need to make sure you have installed onnxruntime and paddleocr (optional for preprocessing utilities).

pip install onnxruntime numpy opencv-python

2. Load the ONNX Models

You need to load both the text detection and recognition models.

import onnxruntime as ort
import numpy as np
import cv2

# Load detection and recognition models
det_model_path = "path_to_detection_model.onnx"
rec_model_path = "path_to_recognition_model.onnx"

det_session = ort.InferenceSession(det_model_path)
rec_session = ort.InferenceSession(rec_model_path)

3. Preprocess Image for Detection

PaddleOCR uses specific preprocessing steps before feeding the image into the model.

def preprocess_image(image_path, target_size=(640,640)):
    image = cv2.imread(image_path)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    original_shape = image.shape[:2]
    
    # Resize and normalize the image
    image = cv2.resize(image, target_size)
    image = image.astype(np.float32) / 255.0
    image = np.transpose(image, (2, 0, 1))  # Convert to CHW format
    image = np.expand_dims(image, axis=0)   # Add batch dimension
    
    return image, original_shape

image_path = "path_to_your_image.jpg"
image, original_shape = preprocess_image(image_path)

4. Perform Text Detection

Run the detection model and apply post-processing.

input_name = det_session.get_inputs()[0].name
outputs = det_session.run(None, {input_name: image})

# The output here needs to be processed to extract bounding boxes
# The processing logic depends on the specific PaddleOCR model used

5. Adjust det_db_unclip_ratio

The det_db_unclip_ratio is a post-processing parameter that controls how much the bounding boxes are expanded. You would generally modify this while processing the detection results.

def adjust_bounding_boxes(boxes, unclip_ratio=1.5):
    # Adjust the bounding boxes by expanding them
    expanded_boxes = []
    for box in boxes:
        x_min, y_min, x_max, y_max = box
        width = x_max - x_min
        height = y_max - y_min
        new_x_min = x_min - (width * (unclip_ratio - 1) / 2)
        new_y_min = y_min - (height * (unclip_ratio - 1) / 2)
        new_x_max = x_max + (width * (unclip_ratio - 1) / 2)
        new_y_max = y_max + (height * (unclip_ratio - 1) / 2)
        expanded_boxes.append([new_x_min, new_y_min, new_x_max, new_y_max])
    return expanded_boxes

After extracting bounding boxes, you can apply this function.

det_db_unclip_ratio = 1.5  # Modify this value as needed
expanded_boxes = adjust_bounding_boxes(detected_boxes, det_db_unclip_ratio)

6. Perform Recognition on Cropped Text Regions

For each detected text region, extract and preprocess it before passing it to the recognition model.

def preprocess_for_recognition(cropped_image):
    cropped_image = cv2.resize(cropped_image, (100, 32))
    cropped_image = cropped_image.astype(np.float32) / 255.0
    cropped_image = np.transpose(cropped_image, (2, 0, 1))
    cropped_image = np.expand_dims(cropped_image, axis=0)
    return cropped_image

for box in expanded_boxes:
    xmin, ymin, xmax, ymax = map(int, box)
    cropped_img = image[:, ymin:ymax, xmin:xmax]  # Crop from image
    
    processed_crop = preprocess_for_recognition(cropped_img)
    input_name = rec_session.get_inputs()[0].name
    rec_output = rec_session.run(None, {input_name: processed_crop})

    # The recognition output needs post-processing to get text

---

Summary

• Load ONNX models for detection and recognition.
• Preprocess input images according to PaddleOCR specifications.
• Run text detection model and extract bounding boxes.
• Modify det_db_unclip_ratio in the bounding box expansion step.
• Process detected regions through the recognition model for final text output.

This should help you integrate PaddleOCR using ONNX Runtime into your code. If you run into issues, check out relevant PaddleOCR discussions:
🔗 PaddleOCR Detection Discussion

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.

[sriram-dsl]

please provide the post process code snippet too