Yolo26s Ottoman Printed Line

Model Overview

• Model Type: Line Detection
• Architecture: YOLO26s-seg (Ultralytics)
• Language: Ottoman Turkish
• Font/Style: Printed Naskh
• Task: Segmenting and detecting individual text lines within predefined text regions.

Intended Use

• Primary Use Case: Detecting text lines inside pre-cropped text regions of Ottoman Turkish printed works in the Naskh font.
• Recommendations: For optimal pipeline performance, it is highly recommended to use a separate page-level region detection model first to extract text blocks, and then pass those cropped regions to this line detection model.
• Limitations: While the model might work on full-page images, it was exclusively trained on region-level crops. Full-page inference may yield lower accuracy or miss structural nuances but it works.

Dataset Information

The dataset was aggregated from 9 different works (New Testament (vol 2, 1857 edition), Eşar-ı Ziya, Kozmoğrafya, yahut, İlm-i ahval-i sema, Mecmua-i Fünûn-i Askeriyye, Medhal-i Fıkh, Muahedat Mecmuası, Mir'at-ı Alem, Muhtasar Tarih-i Umumi, and Solakzade Tarihi.).

• Total Pages and Regions: 286 pages containing 1,789 cropped text regions
• Total Lines: 8,115 lines
• Train Split: 1524 regions (6,890 line instances)
• Validation Split: 275 regions (1,225 line instances)

Performance Metrics

Metrics were recorded on the validation set using the best weights (best.pt). Early stopping was triggered at epoch 93.

Segmentation (Mask) Metrics

Metric	Value
Precision (P)	0.977
Recall (R)	0.990
mAP@50	0.990
mAP@50-95	0.865

Bounding Box Metrics

Metric	Value
Precision (P)	0.977
Recall (R)	0.990
mAP@50	0.990
mAP@50-95	0.868

Model Technical Details

• Parameters: 10,365,727
• GFLOPs: 34.1
• Layers: 139 (fused)
• Weight Size: ~23.5 MB
• Hardware Used: NVIDIA L4 (24GB VRAM) / CUDA:0
• Training Time: ~1.638 hour

Training Configuration

The model was trained for 122 epochs (stopped early via a patience of 20) with a target of 150 epochs.

Key Hyperparameters

• Image Size: 1280
• Batch Size: 16
• Optimizer Learning Rate Scheduling: Cosine LR (cos_lr=True)
• Early Stopping Patience: 20 epochs

Data Augmentations

Augmentations were applied.

• Rotation: 2 degrees
• Translation: 5% (translate=0.05)
• Scaling: 10% (scale=0.1)
• HSV Augmentation: Hue: 0.01, Saturation: 0.2, Value: 0.2
• Disabled Augmentations: Mosaic (0.0), Up-Down Flip (0.0), Left-Right Flip (0.0), Erasing (0.0)

Usage Example

pip install ultralytics       # You should write this as !pip install ultralytics if you use Google Colab.

from ultralytics import YOLO

# Load the trained model
model = YOLO('path/to/yolo26s_line_ota-printed-naskh.pt')

# Run inference on a cropped text region
results = model.predict(source='path/to/region_crop.jpg', imgsz=1280, conf=0.5)

# View results
for r in results:
    r.show()  # Display image with predicted line masks

License

Commercial use is forbidden.

Contact

enes.yilandiloglu@helsinki.fi