A collection of information about Data Augmentation libraries for Image Processing.

KerasCV

last updated on 10/10/2023

Repository Documentation Transformations
Supported languages:
  • Python (3.7+)
Supported ML libraries:
  • Keras
KerasCV
KerasCV
Bounding box formats and utilities
Bounding box formats and utilities
Models
Models
Layers
Layers
Augmentation layers
Augmentation layers
Preprocessing layers
Preprocessing layers
Regularization layers
Regularization layersText is not SVG - cannot display
Transformations Description Tags
1.

Performs the AugMix data augmentation technique.
Original Augmented
Original Augmented 
keras_cv.layers.AugMix(
  value_range=(0, 255),
  severity=0.3,
  num_chains=3,
  chain_depth=[1, 3],
  alpha=1.0,
  seed=None
)
2.

Performs the AutoContrast operation on the image.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.AutoContrast([0,255])
3.

Shuffle channels of the input image.

 PHOTOMETRY
Original Augmented
Original Augmented 
keras_cv.layers.ChannelShuffle(groups=3, seed=None)
4.

Implements the CutMix data augmentation technique.

 MIXED SAMPLE
Original Augmented
Original Augmented 
# images
img_1 = tf.keras.utils.load_img(path + '024.jpg')
img_2 = tf.keras.utils.load_img(path + '025.jpg')

img_1 = tf.keras.utils.img_to_array(img_1)
img_2 = tf.keras.utils.img_to_array(img_2)

images = np.array([img_1, img_2])
labels = np.array([[0], [1]])

labels = tf.one_hot(labels.squeeze(), 2)

# aug
cutmix = keras_cv.layers.preprocessing.cut_mix.CutMix(10)
output = cutmix({"images": images[:2], "labels": labels[:2]})
5.

Performs histogram equalization on a channel-wise basis.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.Equalization(value_range=(0, 255), bins=256)
6.

Implements the FMix data augmentation technique.

 MIXED SAMPLE
Original Augmented
Original Augmented 
# images
img_1 = tf.keras.utils.load_img(path + '024.jpg')
img_2 = tf.keras.utils.load_img(path + '025.jpg')

img_1 = tf.keras.utils.img_to_array(img_1)
img_2 = tf.keras.utils.img_to_array(img_2)

images = np.array([img_1, img_2])
labels = np.array([[0], [1]])

labels = tf.one_hot(labels.squeeze(), 2)

# aug
fourier_mix = keras_cv.layers.preprocessing.FourierMix(0.5)
output = fourier_mix({"images": images[:2], "labels": labels[:2]})
7.

Transforms RGB images to Grayscale images.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.Grayscale(output_channels=1)
8.

Implements GridMask augmentation.

 DROPOUT
Original Augmented 
keras_cv.layers.GridMask(
  ratio_factor=(0, 0.5),
  rotation_factor=0.15,
  fill_mode="constant",
  fill_value=200,
  seed=None,
)
9.

Implements resize with scale distortion.

 BASIC GEOMETRY
Original Augmented 
keras_cv.layers.JitteredResize(
  target_size=(1000,800),
  scale_factor=(0.8, 1.25),
  crop_size=None,
  bounding_box_format=None,
  interpolation="bilinear",
  seed=None
)
10.

Performs the MixUp data augmentation technique. References: “mixup: Beyond Empirical Risk Minimization”, “Bag of Freebies for Training Object Detection Neural Networks”.

 MIXED SAMPLE
Original Augmented
Original Augmented 
# images
img_1 = tf.keras.utils.load_img(path + '024.jpg')
img_2 = tf.keras.utils.load_img(path + '025.jpg')

img_1 = tf.keras.utils.img_to_array(img_1)
img_2 = tf.keras.utils.img_to_array(img_2)

images = np.array([img_1, img_2])
labels = np.array([[0], [1]])

labels = tf.one_hot(labels.squeeze(), 2)

# aug
mixup = keras_cv.layers.preprocessing.MixUp(10)
output = mixup({"images": images[:2], "labels": labels[:2]})
11.

Reduces the number of bits for each color channel. References: “AutoAugment”, “RandAugment”.

 STYLE FILTER
Original Augmented 
keras_cv.layers.Posterization(value_range=(0, 255), 4)
12.

Performs the Rand Augment operation on the input image.
Original Augmented
Original Augmented 
keras_cv.layers.RandAugment(
  value_range=(0, 255),
  augmentations_per_image=3,
  magnitude=0.5,
  magnitude_stddev=0.15,
  rate=0.9090909090909091,
  geometric=True,
  seed=None
)
13.

Constructs a pipeline based on provided arguments.
14.

Randomly shifts values for each channel of the input image(s).

 PHOTOMETRY
Original Augmented
Original Augmented 
keras_cv.layers.RandomChannelShift(value_range=[0, 255], factor=(0.2, 0.2), channels=3, seed=None)
15.

Randomly performs the color degeneration operation. The sharpness operation first converts an image to gray scale, then back to color. It then takes a weighted average between original image and the degenerated image.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.RandomColorDegeneration(factor=(0.4, 0.4), seed=None)
16.

Randomly cuts out rectangles from images and fills them.

 DROPOUT
Original Augmented 
keras_cv.layers.RandomCutout(height_factor=0.1, width_factor=0.3, fill_mode="constant",
                             fill_value=200, seed=None)
Original Augmented 
keras_cv.layers.RandomCutout(height_factor=0.1, width_factor=0.3, fill_mode='gaussian_noise',
                             fill_value=0.0, seed=None)
17.

Randomly adjusts the hue.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.RandomHue(factor=0.3, value_range=[0,255], seed=None)
18.

Randomly adjusts the saturation.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.RandomSaturation(factor=(0.4, 0.4), seed=None)
Original Augmented 
keras_cv.layers.RandomSaturation(factor=(0.7, 0.7), seed=None)
19.

Randomly performs the sharpness operation. This operation first performs a blur operation, then blends between the original image and the blurred image. It makes the edges of an image less sharp than they were in the original image. Reference: ImageEnhance.

 SHARPNESS
Original Augmented 
keras_cv.layers.RandomSharpness(factor=(1, 1), value_range=[0, 255], seed=None)
20.

A preprocessing layer which randomly shears images during training. This layer will apply random shearings to each image, filling empty space according to fill_mode.

 BASIC GEOMETRY
Original Augmented 
keras_cv.layers.RandomShear(
  x_factor=(0.3, 0.3),
  y_factor=None,
  interpolation="bilinear",
  fill_mode="reflect",
  fill_value=0.0,
  bounding_box_format=None,
  seed=None
)
Original Augmented 
keras_cv.layers.RandomShear(
  x_factor=None,
  y_factor=0.4,
  interpolation="bilinear",
  fill_mode="wrap",
  fill_value=0.0,
  bounding_box_format=None,
  seed=None
)
21.

A preprocessing layer which resizes the image.

 BASIC GEOMETRY
Original Augmented 
keras_cv.layers.Resizing(
  height=300,
  width=400,
  interpolation="bilinear",
  crop_to_aspect_ratio=False,
  pad_to_aspect_ratio=False,
  bounding_box_format=None,
)
22.

Applies (max_value - pixel + min_value) for each pixel in the image. When created without threshold parameter, the layer performs solarization to all values. When created with specified threshold the layer only augments pixels that are above the threshold value. References: AutoAugment, RandAugment.

 PHOTOMETRY
Original Augmented 
keras_cv.layers.Solarization(value_range=(0,255), addition_factor=0.0, threshold_factor=0.0, seed=None)