Can't extract numbers with single digit

[zikofs]

Hi All,

I’m using Paddle OCR to extract prices from some products price tags, in most cases it give good results but I noticed that it fails to extract prices with single digit

When I edit the image and add a digit next to the first one, the algorithm succeed to extract the price

I’m using Python with the basic code took from PaddleOCR documentation

from paddleocr import PaddleOCR
ocr = PaddleOCR(use_angle_cls=True, lang='fr')
result = ocr.ocr(image_path, cls=True)
for line in result[0]:
    text = line[1][0] 
    position = line[0]
    print(text, position)

Is there an approach or specific configuration that I can use to solve this issue ?

Thanks in advance

[GreatV]

It appears that PaddleOCR is having trouble detecting single-digit numbers in your images. Based on similar issues discussed in the PaddleOCR community, here are some potential solutions to improve detection:

Possible Solutions:

1. Increase Image DPI
   From the related discussion (source), increasing the image resolution (DPI) can improve OCR accuracy. Try preprocessing the image to enhance its quality before passing it to PaddleOCR.

   from PIL import Image
   
   image = Image.open(image_path)
   image = image.resize((image.width * 2, image.height * 2), Image.LANCZOS)  # Upscale image
   image.save("enhanced_image.png")

2. Image Thresholding & Inversion
   In some cases, applying thresholding or inverting the image colors (white text on black background) can help PaddleOCR detect missing numbers. See if inverting the image improves results:

   import cv2
   import numpy as np
   
   img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
   _, img_thresh = cv2.threshold(img, 127, 255, cv2.THRESH_BINARY_INV)  # Invert colors
   cv2.imwrite("processed_image.png", img_thresh)

   Use the new image in your OCR processing.

3. Adjust OCR Threshold Settings
   Modify the detection threshold settings in PaddleOCR to be more sensitive:

   PaddleOCR(det_db_box_thresh=0.4, det_db_thresh=0.2)  # Experiment with lower values

4. Check Text Detection Algorithm
   Ensure you are using an appropriate text detection algorithm such as "DB" with det_db_unclip_ratio=2.0 to improve box detection for small numbers.

5. Train a Custom Model
   If single-digit recognition is not improving with preprocessing techniques, consider fine-tuning a PaddleOCR recognition model on a dataset that includes images with single digits and similar tags.

Next Steps:

• Try these techniques step by step and evaluate whether they help in detecting single-digit prices.
• If the issue persists, consider opening a discussion on the PaddleOCR GitHub Discussions with sample images for further community insights.

Hope this helps! 🚀

---

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

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

[zikofs]

Tried solutions 1, 2, 3, 4, with no luck
Checked parameters like det_algorithm & det_db_unclip_ratio
I've tried already using lower values of threshold parameters, no luck

[zikofs]

Surprisely when I tried the english model lang='en' PaddleOCR was able to exract the number "4"
Anyone facing the same issue, try to use other models or train your custom model