[PPOCR det] First and last character, half is cut off

[PhuongNDVN]

I try to do fine-turning for English alphabets. Text line are detected well except for the first and last characters. Often, the text bounding box is so tight to the characters. Especially the character like "T", only half of the top part "-" is covered in bounding box. How should I change in config to solve this issue.

Assume that If I can solve this issue by training, should I separate detection and recognition of PPOCR? After detection, I increase bounding box, then put the cropped part to OCR again for recognition?

[GreatV]

The issue of the first and last characters being partially cut off in the bounding box when using PPOCR for text detection is likely due to overly tight bounding boxes generated during detection. Here are steps to improve detection and recognition:

Config Adjustments

1. Increase det_db_box_thresh and det_db_unclip_ratio:

   • det_db_box_thresh: Adjust this threshold to ensure better box generation. Try increasing it slightly to allow more complete character coverage.
   • det_db_unclip_ratio: Increasing this ratio will expand the bounding box slightly, capturing more of the characters' surrounding area.

2. Modify det_model_dir:
   If you are fine-tuning the model, ensure that your detection model is properly trained with sufficient varied examples, including cases with characters near the edges.

3. Use use_dilate = True in preprocessing:
   This helps expand detected regions slightly, allowing better coverage.

---

Post-processing Adjustments

If training configurations alone don’t fully resolve the issue, you can consider:

1. Separating Detection and Recognition:

   • Perform detection first.
   • Expand the bounding box manually (e.g., adding padding around detected text).
   • Pass the adjusted region for recognition.

2. Bounding Box Expansion Script (Example in Python):

   import cv2
   def expand_bbox(bbox, img_shape, ratio=0.1):
       x, y, w, h = bbox
       expand_w = int(w * ratio)
       expand_h = int(h * ratio)
       x_new = max(0, x - expand_w)
       y_new = max(0, y - expand_h)
       w_new = min(img_shape[1], x + w + expand_w) - x_new
       h_new = min(img_shape[0], y + h + expand_h) - y_new
       return [x_new, y_new, w_new, h_new]

3. Alternative Models or Augmentation:
   Experimenting with different det_db_box_thresh values and using data augmentation (rotations, distortions) in training may also help improve bounding box robustness.

---

Potential Related Discussions

Though the provided related discussion (Issue with OCR on a Specific Line Using PPOCRv4) is about unrecognized lines and gibberish results rather than bounding box cutoff, you can check it for additional insights:
Issue with OCR on a Specific Line Using PPOCRv4.

Making these adjustments should improve text detection accuracy and ensure characters aren't partially cut off. 🚀

---

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

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

[PhuongNDVN]

I confirmed that when I add more data with "T" character at first and last position of words, detection result is improved. Thanks