Fine-tuning PaddleOCR for vertical Korean text without breaking horizontal recognition

[RunSongham]

Hello

I have Korean license plates with mixed text orientations (see attached image):

• Vertical text: "경기" (2 characters, left side)
• Horizontal text: "37바 4445" (plate number)

Current situation:

• When I use PP-OCRv5 with use_angle_cls=True, the horizontal text is recognized perfectly
• The vertical text "경기" gets garbled/incorrect output
• However, when I crop each character of "경기" individually, recognition works perfectly

My question:
If I fine-tune the recognition model ONLY with vertical text samples like:

• "경기" (region name)
• "서울", "부산", "대구" (other region names)
• "영업", "렌터" (vehicle type indicators)

Will this solve the vertical text recognition issue WITHOUT affecting the horizontal text recognition performance?

Specifically:

1. Should I prepare training data as:
   a) Original vertical orientation images?
   b) 90-degree rotated images (as PaddleOCR internally rotates them)?
   c) Both?

2. Will fine-tuning with only vertical text data preserve the model's ability to recognize horizontal text?

3. Or do I need to include BOTH vertical and horizontal text in the fine-tuning dataset to maintain overall performance?

4. Looking at my test results in the image, it seems the detection model should be modified, but is fine-tuning the recognition model the correct approach?

Thank you!