What is the full OCR pipeline?

[thebest132]

How does PaddleOCR know which detection box is associated with which recognition text?

[GreatV]

Full OCR Pipeline in PaddleOCR

PaddleOCR follows a structured pipeline to extract text from images. The pipeline involves the following key steps:

1. Preprocessing: The input image is processed, including resizing and normalization, to make it suitable for detection and recognition models.

2. Text Detection:

   • A text detection model (such as DBNet) is used to locate text regions in the image.
   • The model outputs bounding boxes for detected text regions.

3. Bounding Box Post-Processing:

   • The detected bounding boxes are filtered and refined.
   • Some post-processing techniques, such as non-maximum suppression (NMS), may be used to remove duplicate or overlapping boxes.

4. Text Recognition:

   • Cropped regions based on the detection step are sent to a text recognition model.
   • This model (such as CRNN or Transformer-based models) extracts the textual content from the detected regions.

5. Association of Detection Boxes with Recognized Text:

   • The recognition model processes each detected region independently.
   • The order of detected bounding boxes is typically determined based on a reading order heuristic, which is often left-to-right and top-to-bottom for languages that follow this structure.
   • The recognized text is then mapped to its corresponding detection box.

6. Post-Processing:

   • The OCR system may further refine the text output using techniques like lexicon-based correction or language models.

7. Final Output:

   • The final output consists of the detected bounding boxes and their associated recognized text.

How PaddleOCR Maps Detection Boxes to Recognized Text

PaddleOCR ensures that each piece of recognized text is associated correctly with a corresponding detection box by:

• Keeping track of detected bounding boxes order.
• Extracting and processing text from each detected region one by one.
• Using predefined reading order logic (depending on the language direction).

For more details on the detection model used in PaddleOCR, you can refer to the official documentation:
PaddleOCR Detection Model Training

Let me know if you need further clarification! 🚀

---

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

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