Image Augmentation

(hivenas.utils.image_aug)

Image Augmentation methods.

class ImgAug

Bases: object

Element-wise image augmentation methods, used to preprocess a given dataset.

(most affine transformations used are implemented in tensorflow.keras.preprocessing.image.ImageDataGenerator)

static augment(np_tensor)

Used by ImageDataGenerator’s preprocess_function

Parameters

np_tensor (numpy.array) – rank 3 numpy tensor-respresentation of the data sample

Returns

augmented numpy tensor with all applicable transformations/augmentations

Return type

(numpy.array)

static random_contrast(np_tensor)

Apply random contrast augmentation

Parameters

np_tensor (numpy.array) – rank 3 numpy tensor-respresentation of the data sample

Returns

transformed numpy tensor with random contrast

Return type

(numpy.array)

static random_cutout(np_tensor, cutout_color=127)

Randomly applies cutout augmentation to a given rank 3 tensor as defined in [1]. Defaults to grey cutout

[1] DeVries, T., & Taylor, G. W. (2017). Improved regularization of convolutional neural networks with cutout.

Parameters

• np_tensor (numpy.array) – rank 3 numpy tensor-respresentation of the data sample

• cutout_color (int, optional) – RGB-uniform value of the cutout color (defaults to grey ( 127 ). white ( 255 ) and black ( 0 ) are also valid)

Returns

augmented numpy tensor (with a random cutout)

Return type

numpy.array

static random_saturation(np_tensor)

Apply random saturation augmentation (only works on RGB images, skipped on grayscale datasets)

Parameters

np_tensor (numpy.array) – rank 3 numpy tensor-respresentation of the data sample

Returns

transformed numpy tensor with random saturation

Return type

(numpy.array)