korean-handwriting-rec: korean_PP-OCRv5_mobile_rec fine-tuned on 4.6M Korean handwritten and printed words

[VerifiedIdiot]

Hi all. I fine-tuned the PP-OCRv5 Korean recognition model (korean_PP-OCRv5_mobile_rec) on 4.6M Korean word images of handwriting and printed government forms from AI Hub, and released it as korean-handwriting-rec under Apache-2.0.

It keeps the original input/output and the same 11,945-character dictionary, so any PP-OCRv5 ONNX pipeline can use it by swapping the recognition model. With RapidOCR 3.x:

from rapidocr import RapidOCR

engine = RapidOCR(params={
    "Rec.model_path": "model/rec_korean_hw.onnx",
    "Rec.rec_keys_path": "model/ppocrv5_korean_dict.txt",
})
result = engine("scan.png")

Results on held-out sets (word accuracy, the whole word must match):

Evaluation set	korean_PP-OCRv5_mobile_rec	korean-handwriting-rec
Handwritten words (20K)	67.2%	97.4%
Handwriting by 15 unseen writers (6K words)	56.7%	99.1%
Printed text on 5 unseen government forms (6K words)	88.7%	99.2%
Cursive sentences, CER (1.3K, lower is better)	30.0%	7.8%

• 15 MB ONNX, runs on CPU
• Training code (PaddleOCR2Pytorch conversion, two-stage fine-tuning on Apple MPS): https://github.com/VerifiedIdiot/korean-handwriting-rec
• Download (ONNX + dictionaries): https://github.com/VerifiedIdiot/korean-handwriting-rec/releases/tag/v1.0.0
• Model on Hugging Face: https://huggingface.co/Rebeccaaaaaaaa/korean-handwriting-rec

Thanks to the PaddleOCR and RapidOCR teams for the base model and the runtime. Feedback is welcome.