# Preprocessing

`generate_masks.py` produces the guarded background masks used by the
hand-crafted feature extractor.

Museum photographs often include a backdrop, mount or frame around the
artwork. Removing it helps color/light statistics — but salient-object
segmentation (U^2-Net) applied blindly also eats painted borders, halos and
dark backgrounds that ARE the artwork. So every mask must pass five checks
before it is applied (see the module docstring); otherwise the full image is
used. On the gold set this accepts ~10% of images.

Downstream use (in `features/extract_handcrafted.py`):

- **applied** images: spatial features see the mask's bounding-box crop;
  color/light statistics additionally exclude background pixels inside it.
  Background is never zero-filled — that would create fake edges.
- **rejected** images: full image everywhere.

Outputs: `data/masks/<stem>.png` + `data/masks/verdicts.csv`.