dataAugmentation: Data augmentation: rotating and mirroring images, and...

In imageseg: Deep Learning Models for Image Segmentation

Data augmentation: rotating and mirroring images, and adjusting colors

Description

Rotate and/or mirror images to create augmented training data. Optionally, apply a random shift in brightness, saturation and hue to a certain percentage of the images

Usage

dataAugmentation(
  images,
  subset = NULL,
  rotation_angles = 0,
  flip = FALSE,
  flop = FALSE,
  brightness_shift_lim = c(90, 110),
  saturation_shift_lim = c(95, 105),
  hue_shift_lim = c(80, 120),
  fraction_random_BSH = 0
)

Arguments

images	list. Output of loadImages. List with two items ($info: data frame with information about images, $img: tibble containing magick images)
subset	integer. Indices of images to process. Can be useful for only processing subsets of images (e.g. training images, not test/validation images).
rotation_angles	integer. Angles in which to rotate images using image_rotate)?
flip	logical. mirror along horizontal axis (turn images upside-down using image_flip)?
flop	mirror along vertical axis (switch left and right) using image_flop)?
brightness_shift_lim	numeric. Lower and upper limits for argument brightness in image_modulate
saturation_shift_lim	numeric. Lower and upper limits for argument saturation in image_modulate
hue_shift_lim	numeric. Lower and upper limits for argument hue in image_modulate
fraction_random_BSH	numeric. Fraction of images to apply random brightness / saturation / hue shifts to (between 0 and 1)

Details

For creating training data for canopy, rotation and mirroring in both axes is appropriate. For understory vegetation density, only flop images (don't flip), and don't apply a hue shift since recognition of the orange flysheet is color-critical.

Value

A list with 2 elements: $info, a data frame with information about the images, and $img, a tibble with magick images

Examples

# Example 1: Canopy
wd_images_can <- system.file("images/canopy/resized",
                             package = "imageseg")

images_can <- loadImages(imageDir = wd_images_can)

images_can_aug <- dataAugmentation(images = images_can,
                                   rotation_angles = c(0, 90, 180, 270),
                                   flip = TRUE,
                                   flop = TRUE)
images_can_aug

imageseg documentation built on May 30, 2022, 1:07 a.m.