Classical Augmentation

← Back to Synthetic Data Generation

Data augmentation applies deterministic or random transformations to each training image and its corresponding mask [12]. Because the transformations are applied equally to both image and mask, no new annotation is required.

Geometric transforms:

  • Random horizontal / vertical flip

  • Random rotation (arbitrary angle or restricted to multiples of 90°)

  • Random scaling (zoom in / out)

  • Random translation (shift)

  • Elastic deformation: locally random displacement fields that simulate tissue or material deformation

  • Perspective / affine warping

Intensity transforms:

  • Random brightness and contrast adjustment

  • Random gamma correction

  • Gaussian noise addition

  • Random blur (Gaussian, motion)

  • Cutout / random erasing: randomly masking rectangular regions to simulate occlusion

Libraries such as albumentations and torchvision.transforms provide high-performance, composable augmentation pipelines.

Rotating, resizing, or shearing an image exposes empty background at its edges — space the original image didn’t cover. Every JuSPICE method that augments images this way (Aug, PB, PB_NonGauss, SDiff, DCGAN) fills that empty space with IOPaint (the LaMa model), run as a separate program, so the result has no blank border. IOPaint is a separate command-line tool, not one of JuSPICE’s own declared dependencies — see the IOPaint installation page for how to set it up.

See also