paddleocr-ppocr6-mek-finetuned

PP-OCRv6 Small – Magyar Antikva OCR (MEK Fine-tune)

Fine-tuned text recognition model based on PP-OCRv6 small (PP-OCRv6_small_rec, ~5.2M parameters), specialized for Hungarian antiqua typography from the 1880–1930 period, with emphasis on correct recognition of Hungarian-specific characters: Ő, Ű, É, Í (uppercase accented letters).

Model Details

Training Data

Dataset: GaborMadarasz/synthetic_OCR_dataset_MEK

Synthetic text-line images generated from digitized books of the Magyar Elektronikus Könyvtár (MEK), covering Hungarian antiqua typeset publications from approximately 1880 to 1930.

Key characteristics:

• Hungarian language, historical antiqua typography
• Emphasis on accented uppercase characters (Ő, Ű, É, Í) commonly misrecognized by general-purpose OCR models
• Synthetic generation pipeline based on MEK source texts
• Train / validation splits included

Training Configuration

Exported Model Format

This model is distributed as a PaddlePaddle static inference model. The PP-OCRv6 rec architecture exports to a combined format (no separate .pdmodel file):

├── config.json          # HuggingFace metadata
├── inference.json       # Character dictionary (vocabulary)
├── inference.pdiparams  # Model weights + graph (combined static format)
├── inference.yml        # PaddleX pipeline config
└── README.md

Usage

Requirements

pip install paddlepaddle-gpu paddleocr>=3.0.3 huggingface_hub

Inference

# ============================================================
# Inference – GaborMadarasz/paddleocr-ppocr6-mek-finetuned
# ============================================================
import os
import numpy as np
import cv2
from huggingface_hub import snapshot_download
from paddleocr import TextRecognition

# ------------------------------------------------------------
# 1. Download model from HuggingFace
# ------------------------------------------------------------
MODEL_DIR = snapshot_download(
    repo_id="GaborMadarasz/paddleocr-ppocr6-mek-finetuned",
    repo_type="model",
)
print(f"Model downloaded to: {MODEL_DIR}")

# ------------------------------------------------------------
# 2. Load model
# ------------------------------------------------------------
model = TextRecognition(
    model_name="PP-OCRv6_small_rec",
    model_dir=MODEL_DIR,
    engine="paddle",   # required for static graph (.pdiparams) format
    device="gpu",      # or "cpu"
)

# ------------------------------------------------------------
# 3. Smoke test – synthetic Hungarian text image
# ------------------------------------------------------------
def create_smoke_test_image(
    text: str = "Árvíztűrő tükörfúrógép ŐŰÉÍ",
    save_path: str = "/tmp/smoke_test.jpg",
) -> str:
    """Render a text-line image with PIL for smoke testing."""
    from PIL import Image, ImageDraw, ImageFont

    img_h, img_w = 48, 640
    img = Image.new("RGB", (img_w, img_h), color=(255, 255, 255))
    draw = ImageDraw.Draw(img)

    # Try to load a system font; fall back to default if unavailable
    font = None
    for font_path in [
        "/usr/share/fonts/truetype/dejavu/DejaVuSerif.ttf",
        "/usr/share/fonts/truetype/liberation/LiberationSerif-Regular.ttf",
        "/usr/share/fonts/truetype/freefont/FreeSerif.ttf",
    ]:
        if os.path.exists(font_path):
            font = ImageFont.truetype(font_path, size=32)
            print(f"Font loaded: {font_path}")
            break

    if font is None:
        font = ImageFont.load_default()
        print("Warning: using PIL default font – accented chars may not render correctly")

    draw.text((8, 4), text, fill=(0, 0, 0), font=font)
    img.save(save_path)
    print(f"Smoke test image saved: {save_path}  (text: '{text}')")
    return save_path

SMOKE_IMG = create_smoke_test_image()

# ------------------------------------------------------------
# 4. Single image inference
# ------------------------------------------------------------
def recognize(img_path: str) -> dict:
    output = model.predict(input=img_path, batch_size=1)
    for res in output:
        return {
            "text":  res.json["res"]["rec_text"],
            "score": res.json["res"]["rec_score"],
            "image": img_path,
        }

# ------------------------------------------------------------
# 5. Batch inference
# ------------------------------------------------------------
def recognize_batch(img_paths: list, batch_size: int = 8) -> list:
    results = []
    output = model.predict(input=img_paths, batch_size=batch_size)
    for res in output:
        results.append({
            "text":  res.json["res"]["rec_text"],
            "score": res.json["res"]["rec_score"],
            "image": res.json["res"]["input_path"],
        })
    return results

# ------------------------------------------------------------
# 6. Run smoke test
# ------------------------------------------------------------
result = recognize(SMOKE_IMG)
print("\n=== Smoke test result ===")
print(f"  input : {result['image']}")
print(f"  text  : {result['text']}")
print(f"  score : {result['score']:.4f}")

# ------------------------------------------------------------
# 7. Example usage
# ------------------------------------------------------------
if __name__ == "__main__":
    import sys, glob

    img_paths = sys.argv[1:] if len(sys.argv) > 1 else []

    if not img_paths:
        print("\nUsage: python infer.py image1.jpg image2.jpg ...")
        print("No images provided – smoke test result above is the output.")
        sys.exit(0)

    if len(img_paths) == 1:
        r = recognize(img_paths[0])
        print(f"\n[{r['image']}]")
        print(f"  text  : {r['text']}")
        print(f"  score : {r['score']:.4f}")
    else:
        results = recognize_batch(img_paths)
        for r in results:
            print(f"[{r['image']}]  →  \"{r['text']}\"  (score: {r['score']:.4f})")

Note: The model expects pre-cropped single text-line images. For full-page OCR, use a text detection model upstream (e.g. PP-OCRv6_server_det) to detect and crop individual text lines first.

Evaluation

Results on the held-out validation set (MEK synthetic):

End-to-end CER evaluation on the annotated 25-page MEK baseline will be published here once available.

Limitations

• Optimized for Hungarian antiqua typography (1880–1930); performance on modern fonts or other scripts is not guaranteed.
• Text-line input only — not a full-page OCR model.
• The engine="paddle" parameter is required; the transformers and paddle_dynamic engines are not supported without additional weight conversion.

Citation

If you use this model, please cite the base architecture:

@misc{ppocr2020,
  title  = {PaddleOCR: Awesome multilingual OCR toolkits based on PaddlePaddle},
  author = {PaddlePaddle Authors},
  year   = {2020},
  url    = {https://github.com/PaddlePaddle/PaddleOCR}
}

Author

Gábor Madarász — Hungarian NLP hobby project.
Part of a broader effort to improve OCR and NLP tooling for the Hungarian language.
HuggingFace: GaborMadarasz