hksorensen/diagram-detector-model

Diagram Detector - v5

High-accuracy diagram detection model for academic papers using YOLO11 architecture.

Model Description

This model detects diagrams, figures, and illustrations in academic papers. It is trained on mathematical publications with rigorous augmentation to preserve semantic meaning of diagrams.

Performance

Page-Level Metrics (Binary Classification)

• Binary F1: 98.49%
• Precision: 98.58%
• Recall: 98.40%
• Accuracy: 98.37%

Box-Level Metrics (Object Detection)

• mAP50: N/A
• mAP50-95: N/A
• Precision: N/A
• Recall: N/A
• Best Epoch: N/A

Optimal Inference Parameters

• Confidence Threshold: 0.20
• IOU Threshold: 0.30

These thresholds were determined through grid search optimization for best page-level performance.

Usage

from diagram_detector import DiagramDetector

# Initialize detector
detector = DiagramDetector(model='diagram-detector-v5')

# Detect diagrams in a PDF
results = detector.detect_pdf('paper.pdf')

# Or detect in images
results = detector.detect('path/to/images/')

Training Details

Dataset Splits

• Training: 70.00%
• Validation: 15.00%
• Test: 15.00%

Augmentation

• Rotation: ±0.0°
• Translation: ±10%
• Scale: ±20%
• Horizontal/Vertical Flips: Disabled (preserves semantic meaning)

Hyperparameters

• Image Size: 640px
• Batch Size: auto
• Epochs: 200 (patience: 50)
• Optimizer: auto
• Learning Rate: 0.0100 → 0.0100

Limitations

• Optimized for academic papers and mathematical publications
• May require fine-tuning for other document types
• Performance may vary on hand-drawn or low-quality diagrams
• No horizontal/vertical flips during training (maintains orientation semantics)

Citation

@software{diagram_detector_yolo11n_2026,
  author = {Sørensen, Henrik Kragh},
  title = {Diagram Detector Model v5},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/hksorensen/diagram-detector-model}
}

Additional Resources

• Package: pip install diagram-detector
• Repository: Private (contact author for access)

---

Model uploaded: 2026-01-12T19:28:05.111806