Data Augmentation

Apply random transformations to training patches to improve generalization.

Data augmentation artificially expands the training set by applying random transformations to patches during training. It is a checkbox under Advanced Options, and the presets set it differently for each branch.

Preset defaults

• N2V branch: on in every preset

• StructN2V branch: off in every preset

The StructN2V branch keeps augmentation off on purpose: rotating or flipping a patch would rotate the directional noise structure too, breaking the alignment between the patch and its discovered structural mask.

Transformations applied

• Random rotations (90°, 180°, 270°)

• Horizontal and vertical flips

Benefits

• Helps the model generalize to different orientations

• Reduces overfitting on small datasets

Recommendations

• Leave augmentation on for the N2V branch

• Leave it off for the StructN2V branch, matching the presets

Note: Augmentation does not affect the noise measurement or mask discovery in autoStructN2V, which happen on the raw stack before training.