[PPOCR rec] Image size of Fineturning & Inference for custom dataset

[PhuongNDVN]

I am trying to fine-tune PPOCR recognition. If I fix the height of the images used for training at 48, the width of the images ranges from 376 to 460 while maintaining the original aspect ratio. What should I set as the image size for PPOCR recognition during training?

My second question is: should I resize the cropped text-region images while keeping the same aspect ratio before passing them to OCR for inference?

The model I fine-tuned predicts well on my custom data, but the minimum prediction score for individual characters in a word is low.

[GreatV]

For your questions regarding image size during fine-tuning and inference with PPOCR recognition:

1. What should I set as the image size for PPOCR recognition during training?

• Since you fixed the height of images at 48 pixels, and the width varies from 376 to 460 pixels, you should resize the images to a uniform width while keeping the aspect ratio as close as possible.
• From the referenced discussion, the training configuration uses:
  • RecConAug with image_shape: [48, 320, 3]
  • RecResizeImg with image_shape: [3, 48, 320]
• Recommended approach: Consider rescaling your images to a standard width used in PPOCR training, such as 320 pixels, while maintaining the height at 48 pixels. This ensures proper feature extraction and consistency during training.

2. Should I resize cropped text-region images while keeping the aspect ratio before passing them to OCR for inference?

• Yes, but you should ensure that the model receives inputs of a consistent shape. PPOCR models expect input images of a fixed height (e.g., 48 pixels), while the width can vary.
• Suggested approach:
  • Resize the cropped image height to 48 pixels (to match training conditions).
  • Keep the width flexible but pad or resize it to a fixed size such as 320 pixels (as done in training) to avoid distortion.
  • Resizing while maintaining the aspect ratio ensures better recognition accuracy while preventing unwanted artifacts.

3. Low Minimum Prediction Score for Individual Characters

• Your model performs well, but low prediction scores for individual characters could be due to:
  • Incorrect fine-tuning weights: Ensure that pre-trained weights are correctly loaded.
  • Character dictionary issues: If you modified the character set, ensure that the dictionary used matches the training dataset.
  • Text augmentation: Excessive distortions or inconsistent image preprocessing could make it harder for the model to recognize characters accurately.

To diagnose this issue further, consider visualizing low-confidence predictions and fine-tuning thresholds for confidence scores.

For more details, refer to this related discussion on fine-tuning PPOCRv4 recognition:
Discussion: Fine-tuning PPOCRv4 recognition

---

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

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