Image augmentation pipeline

[JLrumberger]

Section 1: Design details

We want to have an image augmentation pipeline that randomly transforms the images. PR #9

Relevant background

DL models are known to become more robust and overfit less when augmentations are used during training. Thus we want to have an augmentation pipeline that applies spatial and intensity augmentations onto images and solely spatial augmentations on the masks.

Design overview

• get_augmentation_pipeline(params) function that takes in a parameter dict and returns an object of class imagaug.augmenters.Sequential.
• augment_images(images, masks, augmentation_pipeline) function that takes in the augmentation pipeline, input images and masks and returns augmented images and masks

Design list/flowchart

• get_augmentation_pipeline(params) produces pipeline object
• augment_images(images, masks, augmentation_pipeline) applies augmentation pipeline onto our data. For images bilinear interpolation is used, for masks nearest-neighbor interpolation so that their unique values don't change.

Required inputs

• `params = {

  flip

  "flip_prob": 1.0,
  # affine
  "affine_prob": 1.0, "scale_min": 0.5, "scale_max": 1.5, "shear_angle": 10,
  # elastic
  "elastic_prob": 1.0, "elastic_alpha": [0, 5.0], "elastic_sigma": 0.5,
  # rotate
  "rotate_count": [0, 3],
  # gaussian noise
  "gaussian_noise_prob": 1.0, "gaussian_noise_min": 0.1, "gaussian_noise_max": 0.5,
  # gaussian blur
  "gaussian_blur_prob": 0.0, "gaussian_blur_min": 0.1, "gaussian_blur_max": 0.5,
  # contrast aug
  "contrast_prob": 0.0, "contrast_min": 0.1, "contrast_max": 2.0,

  }`

• images (np.array, dtype=float32) B x H x W x C
• masks (np.array, dtype=uint8/16) B x H x W x C

Output files

Intermediate output: augmentation pipeline object of class imagaug.augmenters.Sequential Final output: augmented images (np.array, dtype=float32) B x H x W x C and masks (np.array, dtype=uint8/16) B x H x W x C

Section 2: Implementation details

Control flow

Images need to be interpolated with bilinear interpolation, masks need to be interpolated with nearest-neighbor interpolation, otherwise the they will contain label values that weren't there before.

Milestones and timeline

Ready for review.