5️⃣ Improvements in Data Augmentation

[Microcosmos22]

Forgery areas can be tiny. Strong spatial augmentation can
→ rotate/move them out of image bounds
→ distort mask connections
We should add:

ShiftScaleRotate with smaller limits

GaussianBlur / JPEG compression (relevant for forensic artifacts)

Cutout or PatchDropout (combat annotation gaps)

[Microcosmos22]

1. Find out how many of the forged images are:

• Copy-move forgeries: Check the masks, and if any two regions are similar (rotation, scaling)
• Deletion forgeries: Regions are not similar (any two pairs are not similar)

2. Replicate forgeries and augment the dataset.

[Microcosmos22]

Current data count is very balanced:

Run data_augmentation.py

forged 2723, non-forged: 2405
copy_paste: 1730

[Microcosmos22]

Could create new forgeries either from using the copy-pasted ones and easily moving one of the mask*images.

Another way is to grab the authentic images, detect and object then copy-paste-move it.