Strikethrough Detection Model (YOLOv8m)

A YOLOv8m object detection model fine-tuned to detect strikethrough marks on handwritten math documents.

Training Details

• Architecture: YOLOv8m (25.8M parameters)
• Training data: 20,000 synthetic document pages with varied strikethrough styles
• Hardware: 4x NVIDIA H100 80GB (Modal)
• Epochs: 200
• Image size: 1280px
• Optimizer: AdamW (lr=0.001, cosine schedule)

Performance (Test Set)

Metric	Value
Precision	0.9963
Recall	0.9897
mAP@50	0.9945
mAP@50-95	0.9799

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("document.png", conf=0.5)

Strikethrough Types Detected

• Single horizontal lines
• Cross-out X marks
• Scribble overlays
• Full-page X marks
• Multi-expression cross-outs
• Region scribbles
• Horizontal sweeps
• Dense page scribbles