This project implements hyperbola extraction on GPR patch images: a ResNet50-based segmentation model that splits hyperbola (foreground) from background, similar to the two-part split used in the /resnet folder (e.g. best_on_folder7).

• Backbone: ResNet50 (1‑channel input, pretrained on ImageNet by default).
• Head: Decoder producing a binary mask (hyperbola vs background) per patch.
• Input: PNG GPR patches named like *_patch*.png with YOLO‑style label files *_patch*.txt. Labels are converted to ellipse masks inside the bounding box for training.
• Output:
  • Trained checkpoints under output_dir/exp_<id>/fold_k/.
  • Segmentation mask (extracted hyperbola region); optional visualizations: raw patch, GT mask, predicted mask.

The command‑line interface mirrors the existing /resnet project so you can reuse scripts and habits with minimal changes.

See INSTALLATION.md for environment setup, COMMANDS.md for ready‑to‑run commands, and MANUAL.md for a full workflow description.

These are rough targets for reasonable extraction (segmentation) (numbers will vary by dataset and split):

If train_iou and val_iou climb into the 0.5–0.8 range without val_loss diverging, the model is usually extracting the hyperbola region reliably.