Missing first line of paragraph

[guptakvgaurav]

🔎 Search before asking

• I have searched the PaddleOCR Docs and found no similar bug report.
• I have searched the PaddleOCR Issues and found no similar bug report.
• I have searched the PaddleOCR Discussions and found no similar bug report.

🐛 Bug (问题描述)

I am using this OCR on insurance reimbursement policy document. I found OCR sometime misses the entire line. Whenever it misses the entire line it is usually the first line of the paragraph.

In the attached file, missed lines are in red box:

File with missing text highlighted

Original file (same as above but without highlight)

🏃‍♂️ Environment (运行环境)

OS              macOS-14.4 (23E214)
Environment     Jupyter
Python         3.11.2
PaddleOCR      2.8.1
Install        git
RAM            32.00 GB
CPU            Apple M3 Pro
CUDA           None

🌰 Minimal Reproducible Example (最小可复现问题的Demo)

ocr = PaddleOCR(lang="en", use_angle_cls=True)
image = cv2.imread('temp_page_5.jpg')
result = ocr.ocr(image, cls=True)

[GreatV]

This appears to be a known issue where PaddleOCR may miss lines — often the first line of a paragraph — particularly in structured or dense documents such as reimbursement policy files. Based on your description and images, the red-marked areas clearly show missing lines that should have been detected and recognized.

The potential causes for this behavior include:

1. Text Detection Issues:
   PaddleOCR uses a text detection model (by default, DBNet) to locate text regions before recognition. If the first line is very close to the edge of the document or has less visual contrast, it might not be detected. This is especially likely if the document has a narrow margin, tightly packed lines, or a header section that isn’t clearly distinguishable from the body.

2. Preprocessing or Cropping:
   The image might be undergoing resizing or padding during pre-processing, which could slightly crop out or distort the top part of the image, affecting the ability to detect the first line.

3. Detector Parameters:
   The default detection configuration may not be optimal for your type of document. Particularly:

   • det_db_box_thresh
   • det_db_thresh
   • det_db_unclip_ratio
     These thresholds can significantly affect which boxes are retained. A setting that is too high may suppress less confidently detected lines — such as a faint header line.

4. Angle Classification (use_angle_cls=True):
   In some cases, auto angle classification might slightly rotate the image in a way that shifts the top line out of view or makes spacing inconsistent.

🛠 Recommended Steps to Mitigate:

• Try adjusting detection parameters using a custom inference script. In particular:

  ocr = PaddleOCR(lang="en", use_angle_cls=True,
                  det_db_box_thresh=0.3,  # lower this to detect weaker boxes
                  det_db_thresh=0.2,
                  det_db_unclip_ratio=2.0)

• Check if the missed lines are very close to page margins. If so, consider padding the image manually before feeding it into OCR:

  import cv2
  image = cv2.imread('temp_page_5.jpg')
  padded = cv2.copyMakeBorder(image, 50, 0, 0, 0, cv2.BORDER_CONSTANT, value=(255,255,255))
  result = ocr.ocr(padded, cls=True)

• Confirm that the image is not being auto-cropped or resized in a way that could remove important content at the top.

• Try upgrading to the latest version of PaddleOCR or testing with a different detection backend (e.g., CRAFT or EAST) to compare behavior.

📌 Related Reference:
This issue appears similar to Issue #14275, where some content (e.g., "工程记录") is not getting recognized at all. See here: #14275

Though that example is in Chinese and not specific to missing first lines, the underlying cause — missed detection — is relevant.

Let me know if you’d like help adjusting detection thresholds or running a debug visualization to confirm whether the detector is missing the lines entirely.

---

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

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