Macedonian OCR V8

Fine-tuned PaddleOCR recognition model for Macedonian Cyrillic text in scanned books, paired with PP-OCRv6 medium detection.

Results

Engine	CER
V8+LM (this model + KenLM post-processing)	0.20%
V8 (this model, raw)	0.26%
V7 (previous)	1.02%
Tesseract (mkd)	1.43%
Qwen 3.5 122B (best VLM)	0.39%

Measured on 10 held-out scanned Macedonian book pages.

Usage

from paddleocr import PaddleOCR
ocr = PaddleOCR(
    text_detection_model_name='PP-OCRv6_medium_det',
    text_recognition_model_name='PP-OCRv5_server_rec',
    text_recognition_model_dir='./mk_rec_v8_infer',
    text_det_thresh=0.15,
    text_det_box_thresh=0.28,
    text_det_unclip_ratio=3.0,
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False,
)
result = ocr.predict('your_page.jpg')
for page in result:
    for text, score in zip(page['rec_texts'], page['rec_scores']):
        print(f'{score:.2f}  {text}')

Download

pip install huggingface_hub
python3 -c "
from huggingface_hub import snapshot_download
snapshot_download('mjurukov/macedonian-ocr-v8', local_dir='./mk_rec_v8_infer')
"

Training

Base model: PP-OCRv5 mobile SVTR_LCNet with MultiHead (CTC + NRTR)
Dictionary: 170 characters — full Macedonian Cyrillic, Latin, digits, punctuation
Detection: PP-OCRv6_medium_det (34.5M params) — 36% faster than v5
Training data: 583k synthetic + 172k targeted + 42k IED-boost + 1,674 real annotated lines
Post-processing: KenLM 3-gram language model from 1,065 Macedonian books + 256k-word frequency dictionary
Hardware: NVIDIA RTX 2080 Ti (11 GB), 50 epochs, ~4 days training

Preprocessing

Winning pipeline (validated via A/B testing):

Grayscale
Deskew (HoughLinesP, max ±15°)
Illumination flattening (morphological close divide)
Unsharp masking (gain 0.4, σ=1.5)

License

Model weights: CC BY 4.0 — free to use, modify, and redistribute with attribution. Code: MIT

mjurukov
/

macedonian-ocr-v8