Jawi Manuscript YOLO Text Line Detector

This model is a specialized YOLO-based Text Line Detector fine-tuned for high-precision bounding box extraction from historical and verified Jawi (Arabic script) manuscripts and document folios.

Model Highlights

Architecture: Ultralytics YOLO trained with document layout inductive biases.
Task: Single-class dense text line detection and bounding box localization (line).
Target Script: Jawi (Malay written in Arabic script).
Post-Processing Support: Compatible with Right-to-Left (RTL) baseline reading order sorting and Horizontal Projection Profile (HPP) line splitting.

Usage with Ultralytics

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# Download best weights from Hugging Face Hub
weights_path = hf_hub_download(
    repo_id="Shadowzer0/Jawi_YOLO_Line_Detector",
    filename="best.pt"
)

# Load model and run inference
model = YOLO(weights_path)
results = model.predict("path/to/manuscript_page.jpg", conf=0.25, iou=0.45)

for box in results[0].boxes.xyxy:
    x1, y1, x2, y2 = box.tolist()
    print(f"Detected line: ({x1}, {y1}) -> ({x2}, {y2})")

Integration with Karya Agung Pipeline

This model serves as the primary line segmenter for the Karya Agung Jawi OCR & Transliteration Studio.

Downloads last month: -