About

This python library helps you with augmenting images for your machine learning projects. It converts a set of input images into a new, much larger set of slightly altered images.

Features:

• Most standard augmentation techniques available.
• Techniques can be applied to both images and keypoints/landmarks on images.
• Define your augmentation sequence once at the start of the experiment, then apply it many times.
• Define flexible stochastic ranges for each augmentation, e.g. "rotate each image by a value between -45 and 45 degrees" or "rotate each image by a value sampled from the normal distribution N(0, 5.0)".
• Easily convert all stochastic ranges to deterministic values in order to augment different batches of images in exactly the same way. (E.g. images and their respective heatmaps. If an image is rotated, you want its heatmap to be rotated by the exactly same amount.)
• Optionally run the augmentations in background processes, improving performance of experiments.

The image below shows examples for each available augmentation technique.

Noop is an augmenter that does nothing. Values for crop are (top pixel, right px, bottom px, left px). Other values written in the form (a, b) mean that each value x was randomly picked from the range a <= x <= b.

Requirements and installation

Required packages:

• six
• numpy
• scipy
• scikit-image (pip install -U scikit-image)
• OpenCV (i.e. cv2)

OpenCV has to be manually installed. The other package should auto-install themselves.

To install, simply use sudo pip install imgaug. That version might be outdated though. To always get the newest version directly from github use sudo pip install git+https://github.com/aleju/imgaug. Alternatively, you can download the repository via git clone https://github.com/aleju/imgaug and install by using python setup.py sdist && sudo pip install dist/imgaug-0.2.4.tar.gz.

To deinstall the library, just execute sudo pip uninstall imgaug.

The library is currently only tested in python2.7, but the code is written so that it should run in python3 too.

Examples

A standard machine learning situation. Train on batches of images and augment each batch via crop, horizontal flip ("Fliplr") and gaussian blur:

from imgaug import augmenters as iaa

seq = iaa.Sequential([
    iaa.Crop(px=(0, 16)), # crop images from each side by 0 to 16px (randomly chosen)
    iaa.Fliplr(0.5), # horizontally flip 50% of the images
    iaa.GaussianBlur(sigma=(0, 3.0)) # blur images with a sigma of 0 to 3.0
])

for batch_idx in range(1000):
    # 'images' should be either a 4D numpy array of shape (N, height, width, channels)
    # or a list of 3D numpy arrays, each having shape (height, width, channels).
    # Grayscale images must have shape (height, width, 1) each.
    # All images must have numpy's dtype uint8. Values are expected to be in
    # range 0-255.
    images = load_batch(batch_idx)
    images_aug = seq.augment_images(images)
    train_on_images(images_aug)

Apply heavy augmentations to images (used to create the image at the very top of this readme):

import imgaug as ia
from imgaug import augmenters as iaa
import numpy as np

# random example images
images = np.random.randint(0, 255, (16, 128, 128, 3), dtype=np.uint8)

# Sometimes(0.5, ...) applies the given augmenter in 50% of all cases,
# e.g. Sometimes(0.5, GaussianBlur(0.3)) would blur roughly every second image.
sometimes = lambda aug: iaa.Sometimes(0.5, aug)

# Define our sequence of augmentation steps that will be applied to every image
# All augmenters with per_channel=0.5 will sample one value _per image_
# in 50% of all cases. In all other cases they will sample new values
# _per channel_.
seq = iaa.Sequential(
    [
        # apply the following augmenters to most images
        iaa.Fliplr(0.5), # horizontally flip 50% of all images
        iaa.Flipud(0.2), # vertically flip 20% of all images
        sometimes(iaa.Crop(percent=(0, 0.1))), # crop images by 0-10% of their height/width
        sometimes(iaa.Affine(
            scale={"x": (0.8, 1.2), "y": (0.8, 1.2)}, # scale images to 80-120% of their size, individually per axis
            translate_percent={"x": (-0.2, 0.2), "y": (-0.2, 0.2)}, # translate by -20 to +20 percent (per axis)
            rotate=(-45, 45), # rotate by -45 to +45 degrees
            shear=(-16, 16), # shear by -16 to +16 degrees
            order=[0, 1], # use nearest neighbour or bilinear interpolation (fast)
            cval=(0, 255), # if mode is constant, use a cval between 0 and 255
            mode=ia.ALL # use any of scikit-image's warping modes (see 2nd image from the top for examples)
        )),
        # execute 0 to 5 of the following (less important) augmenters per image
        # don't execute all of them, as that would often be way too strong
        iaa.SomeOf((0, 5),
            [
                sometimes(iaa.Superpixels(p_replace=(0, 1.0), n_segments=(20, 200))), # convert images into their superpixel representation
                iaa.OneOf([
                    iaa.GaussianBlur((0, 3.0)), # blur images with a sigma between 0 and 3.0
                    iaa.AverageBlur(k=(2, 7)), # blur image using local means with kernel sizes between 2 and 7
                    iaa.MedianBlur(k=(3, 11)), # blur image using local medians with kernel sizes between 2 and 7
                ]),
                iaa.Sharpen(alpha=(0, 1.0), lightness=(0.75, 1.5)), # sharpen images
                iaa.Emboss(alpha=(0, 1.0), strength=(0, 2.0)), # emboss images
                # search either for all edges or for directed edges
                sometimes(iaa.OneOf([
                    iaa.EdgeDetect(alpha=(0, 0.7)),
                    iaa.DirectedEdgeDetect(alpha=(0, 0.7), direction=(0.0, 1.0)),
                ])),
                iaa.AdditiveGaussianNoise(loc=0, scale=(0.0, 0.05*255), per_channel=0.5), # add gaussian noise to images
                iaa.OneOf([
                    iaa.Dropout((0.01, 0.1), per_channel=0.5), # randomly remove up to 10% of the pixels
                    iaa.CoarseDropout((0.03, 0.15), size_percent=(0.02, 0.05), per_channel=0.2),
                ]),
                iaa.Invert(0.05, per_channel=True), # invert color channels
                iaa.Add((-10, 10), per_channel=0.5), # change brightness of images (by -10 to 10 of original value)
                iaa.Multiply((0.5, 1.5), per_channel=0.5), # change brightness of images (50-150% of original value)
                iaa.ContrastNormalization((0.5, 2.0), per_channel=0.5), # improve or worsen the contrast
                iaa.Grayscale(alpha=(0.0, 1.0)),
                sometimes(iaa.ElasticTransformation(alpha=(0.5, 3.5), sigma=0.25)), # move pixels locally around (with random strengths)
                sometimes(iaa.PiecewiseAffine(scale=(0.01, 0.05))) # sometimes move parts of the image around
            ],
            random_order=True
        )
    ],
    random_order=True
)

images_aug = seq.augment_images(images)

Quickly show example results of your augmentation sequence:

from imgaug import augmenters as iaa
import numpy as np

images = np.random.randint(0, 255, (16, 128, 128, 3), dtype=np.uint8)
seq = iaa.Sequential([iaa.Fliplr(0.5), iaa.GaussianBlur((0, 3.0))])

# show an image with 8*8 augmented versions of image 0
seq.show_grid(images[0], cols=8, rows=8)

# Show an image with 8*8 augmented versions of image 0 and 8*8 augmented
# versions of image 1. The identical augmentations will be applied to
# image 0 and 1.
seq.show_grid([images[0], images[1]], cols=8, rows=8)

Augment two batches of images in exactly the same way (e.g. horizontally flip 1st, 2nd and 5th images in both batches, but do not alter 3rd and 4th images):

from imgaug import augmenters as iaa

# Standard scenario: You have N RGB-images and additionally 21 heatmaps per image.
# You want to augment each image and its heatmaps identically.
images = np.random.randint(0, 255, (16, 128, 128, 3), dtype=np.uint8)
heatmaps = np.random.randint(0, 255, (16, 128, 128, 21), dtype=np.uint8)

seq = iaa.Sequential([iaa.GaussianBlur((0, 3.0)), iaa.Affine(translate_px={"x": (-40, 40)})])

# Convert the stochastic sequence of augmenters to a deterministic one.
# The deterministic sequence will always apply the exactly same effects to the images.
seq_det = seq.to_deterministic() # call this for each batch again, NOT only once at the start
images_aug = seq_det.augment_images(images)
heatmaps_aug = seq_det.augment_images(heatmaps)

Augment images and landmarks on these images:

import imgaug as ia
from imgaug import augmenters as iaa
from scipy import misc
import random
import numpy as np
images = np.random.randint(0, 50, (4, 128, 128, 3), dtype=np.uint8)

# Generate random keypoints.
# The augmenters expect a list of imgaug.KeypointsOnImage.
keypoints_on_images = []
for image in images:
    height, width = image.shape[0:2]
    keypoints = []
    for _ in range(4):
        x = random.randint(0, width-1)
        y = random.randint(0, height-1)
        keypoints.append(ia.Keypoint(x=x, y=y))
    keypoints_on_images.append(ia.KeypointsOnImage(keypoints, shape=image.shape))

seq = iaa.Sequential([iaa.GaussianBlur((0, 3.0)), iaa.Affine(scale=(0.5, 0.7))])
seq_det = seq.to_deterministic() # call this for each batch again, NOT only once at the start

# augment keypoints and images
images_aug = seq_det.augment_images(images)
keypoints_aug = seq_det.augment_keypoints(keypoints_on_images)

# Example code to show each image and print the new keypoints coordinates
for img_idx, (image_before, image_after, keypoints_before, keypoints_after) in enumerate(zip(images, images_aug, keypoints_on_images, keypoints_aug)):
    image_before = keypoints_before.draw_on_image(image_before)
    image_after = keypoints_after.draw_on_image(image_after)
    misc.imshow(np.concatenate((image_before, image_after), axis=1)) # before and after
    for kp_idx, keypoint in enumerate(keypoints_after.keypoints):
        keypoint_old = keypoints_on_images[img_idx].keypoints[kp_idx]
        x_old, y_old = keypoint_old.x, keypoint_old.y
        x_new, y_new = keypoint.x, keypoint.y
        print("[Keypoints for image #%d] before aug: x=%d y=%d | after aug: x=%d y=%d" % (img_idx, x_old, y_old, x_new, y_new))

Apply single augmentations to images:

from imgaug import augmenters as iaa
import numpy as np
images = np.random.randint(0, 255, (16, 128, 128, 3), dtype=np.uint8)

flipper = iaa.Fliplr(1.0) # always horizontally flip each input image
images[0] = flipper.augment_image(images[0]) # horizontally flip image 0

vflipper = iaa.Flipud(0.9) # vertically flip each input image with 90% probability
images[1] = vflipper.augment_image(images[1]) # probably vertically flip image 1

blurer = iaa.GaussianBlur(3.0)
images[2] = blurer.augment_image(images[2]) # blur image 2 by a sigma of 3.0
images[3] = blurer.augment_image(images[3]) # blur image 3 by a sigma of 3.0 too

translater = iaa.Affine(translate_px={"x": -16}) # move each input image by 16px to the left
images[4] = translater.augment_image(images[4]) # move image 4 to the left

scaler = iaa.Affine(scale={"y": (0.8, 1.2)}) # scale each input image to 80-120% on the y axis
images[5] = scaler.augment_image(images[5]) # scale image 5 by 80-120% on the y axis

Apply an augmenter to only specific image channels:

from imgaug import augmenters as iaa
import numpy as np

# fake RGB images
images = np.random.randint(0, 255, (16, 128, 128, 3), dtype=np.uint8)

# add a random value from the range (-30, 30) to the first two channels of
# input images (e.g. to the R and G channels)
aug = iaa.WithChannels(
  channels=[0, 1],
  children=iaa.Add((-30, 30))
)

images_aug = aug.augment_images(images)

You can use more unusual distributions for the stochastic parameters of each augmenter:

from imgaug import augmenters as iaa
from imgaug import parameters as iap
import numpy as np
images = np.random.randint(0, 255, (16, 128, 128, 3), dtype=np.uint8)

# Blur by a value sigma which is sampled from a uniform distribution
# of range 0.1 <= x < 3.0.
# The convenience shortcut for this is: iaa.GaussianBlur((0.1, 3.0))
blurer = iaa.GaussianBlur(iap.Uniform(0.1, 3.0))
images_aug = blurer.augment_images(images)

# Blur by a value sigma which is sampled from a normal distribution N(1.0, 0.1),
# i.e. sample a value that is usually around 1.0.
# Clip the resulting value so that it never gets below 0.1 or above 3.0.
blurer = iaa.GaussianBlur(iap.Clip(iap.Normal(1.0, 0.1), 0.1, 3.0))
images_aug = blurer.augment_images(images)

# Same again, but this time the mean of the normal distribution is not constant,
# but comes itself from a uniform distribution between 0.5 and 1.5.
blurer = iaa.GaussianBlur(iap.Clip(iap.Normal(iap.Uniform(0.5, 1.5), 0.1), 0.1, 3.0))
images_aug = blurer.augment_images(images)

# Use for sigma one of exactly three allowed values: 0.5, 1.0 or 1.5.
blurer = iaa.GaussianBlur(iap.Choice([0.5, 1.0, 1.5]))
images_aug = blurer.augment_images(images)

# Sample sigma from a discrete uniform distribution of range 1 <= sigma <= 5,
# i.e. sigma will have any of the following values: 1, 2, 3, 4, 5.
blurer = iaa.GaussianBlur(iap.DiscreteUniform(1, 5))
images_aug = blurer.augment_images(images)

You can dynamically deactivate augmenters in an already defined sequence:

import imgaug as ia
from imgaug import augmenters as iaa
import numpy as np

# images and heatmaps, just arrays filled with value 30
images = np.ones((16, 128, 128, 3), dtype=np.uint8) * 30
heatmaps = np.ones((16, 128, 128, 21), dtype=np.uint8) * 30

# add vertical lines to see the effect of flip
images[:, 16:128-16, 120:124, :] = 120
heatmaps[:, 16:128-16, 120:124, :] = 120

seq = iaa.Sequential([
  iaa.Fliplr(0.5, name="Flipper"),
  iaa.GaussianBlur((0, 3.0), name="GaussianBlur"),
  iaa.Dropout(0.02, name="Dropout"),
  iaa.AdditiveGaussianNoise(scale=0.01*255, name="MyLittleNoise"),
  iaa.AdditiveGaussianNoise(loc=32, scale=0.0001*255, name="SomeOtherNoise"),
  iaa.Affine(translate_px={"x": (-40, 40)}, name="Affine")
])

# change the activated augmenters for heatmaps,
# we only want to execute horizontal flip, affine transformation and one of
# the gaussian noises
def activator_heatmaps(images, augmenter, parents, default):
    if augmenter.name in ["GaussianBlur", "Dropout", "MyLittleNoise"]:
        return False
    else:
        # default value for all other augmenters
        return default
hooks_heatmaps = ia.HooksImages(activator=activator_heatmaps)

seq_det = seq.to_deterministic() # call this for each batch again, NOT only once at the start
images_aug = seq_det.augment_images(images)
heatmaps_aug = seq_det.augment_images(heatmaps, hooks=hooks_heatmaps)

Images can be augmented in background processes using the method augment_batches(batches, background=True), where batches is expected to be a list of image batches or a list of batches/lists of imgaug.KeypointsOnImage or a list of imgaug.Batch. The following example augments a list of image batches in the background:

import imgaug as ia
from imgaug import augmenters as iaa
import numpy as np
from skimage import data

# Number of batches and batch size for this example
nb_batches = 10
batch_size = 32

# Example augmentation sequence to run in the background
augseq = iaa.Sequential([
    iaa.Fliplr(0.5),
    iaa.CoarseDropout(p=0.1, size_percent=0.1)
])

# For simplicity, we use the same image here many times
astronaut = data.astronaut()
astronaut = ia.imresize_single_image(astronaut, (64, 64))

# Make batches out of the example image (here: 10 batches, each 32 times
# the example image)
batches = []
for _ in range(nb_batches):
    batches.append(
        np.array(
            [astronaut for _ in range(batch_size)],
            dtype=np.uint8
        )
    )

# Show the augmented images.
# Note that augment_batches() returns a generator.
for images_aug in augseq.augment_batches(batches, background=True):
    misc.imshow(ia.draw_grid(images_aug, cols=8))

Images can also be augmented in background processes using the classes imgaug.BatchLoader and imgaug.BackgroundAugmenter, which offer a bit more flexibility. (augment_batches() is a wrapper around these.) Using these classes is good practice, when you have a lot of images that you don't want to load at the same time.

import imgaug as ia
from imgaug import augmenters as iaa
import numpy as np
from skimage import data

# Example augmentation sequence to run in the background.
augseq = iaa.Sequential([
    iaa.Fliplr(0.5),
    iaa.CoarseDropout(p=0.1, size_percent=0.1)
])

# A generator that loads batches from the hard drive.
def load_batches():
    # Here, load 10 batches of size 4 each.
    # You can also load an infinite amount of batches, if you don't train
    # in epochs.
    batch_size = 4
    nb_batches = 10

    # Here, for simplicity we just always use the same image.
    astronaut = data.astronaut()
    astronaut = ia.imresize_single_image(astronaut, (64, 64))

    for i in range(nb_batches):
        # A list containing all images of the batch.
        batch_images = []
        # A list containing IDs of images in the batch. This is not necessary
        # for the background augmentation and here just used to showcase that
        # you can transfer additional information.
        batch_data = []

        # Add some images to the batch.
        for b in range(batch_size):
            batch_images.append(astronaut)
            batch_data.append((i, b))

        # Create the batch object to send to the background processes.
        batch = ia.Batch(
            images=np.array(batch_images, dtype=np.uint8),
            data=batch_data
        )

        yield batch

# background augmentation consists of two components:
#  (1) BatchLoader, which runs in a Thread and calls repeatedly a user-defined
#      function (here: load_batches) to load batches (optionally with keypoints
#      and additional information) and sends them to a queue of batches.
#  (2) BackgroundAugmenter, which runs several background processes (on other
#      CPU cores). Each process takes batches from the queue defined by (1),
#      augments images/keypoints and sends them to another queue.
# The main process can then read augmented batches from the queue defined
# by (2).
batch_loader = ia.BatchLoader(load_batches)
bg_augmenter = ia.BackgroundAugmenter(batch_loader, augseq)

# Run until load_batches() returns nothing anymore. This also allows infinite
# training.
while True:
    print("Next batch...")
    batch = bg_augmenter.get_batch()
    if batch is None:
        print("Finished epoch.")
        break
    images_aug = batch.images_aug

    print("Image IDs: ", batch.data)

    misc.imshow(np.hstack(list(images_aug)))

batch_loader.terminate()
bg_augmenter.terminate()

List of augmenters

The following is a list of augmenters. Note that most defines variables can be set as ranges, e.g. A=(0.0, 1.0) to sample a random value between 0 and 1.0 per image.

Augmenter	Description
Sequential(C, R)	Takes a list of child augmenters C and applies them in that order to images. If R is true (default: false), then the order is random (chosen once per batch).
SomeOf(N, C, R)	Applies N randomly selected augmenters from from a list of augmenters C to each image. The augmenters are chosen per image. R is the same as for Sequential. N can be a range, e.g. (1, 3) in order to pick 1 to 3.
OneOf(C)	Identical to SomeOf(1, C).
Sometimes(P, C, D)	Augments images with probability P by using child augmenters C, otherwise uses D. D can be None, then only P percent of all images are augmented via C.
InColorspace(T, F, C)	Transforms images from colorspace F (default: RGB) to colorspace T, applies augmenters C and then converts back to F.
WithChannels(H, C)	Selects from each image channels H (e.g. [0,1] for red and green in RGB images), applies child augmenters C to these channels and merges the result back into the original images.
Noop()	Does nothing. (Useful for validation/test.)
Lambda(I, K)	Applies lambda function I to images and K to keypoints.
AssertLambda(I, K)	Checks images via lambda function I and keypoints via K and raises an error if false is returned by either of them.
AssertShape(S)	Raises an error if input images are not of shape S.
Crop(PX, PC, KS)	Cuts away either PX pixels or PC percent of pixels at top/right/bottom/left of images. If KS is true (default), the resulting image is resized back to the original size.
Fliplr(P)	Horizontally flips images with probability P.
Flipud(P)	Vertically flips images with probability P.
Superpixels(P, N, M)	Generates N superpixels of the image at (max) resolution M and resizes back to the original size. Then P percent of all superpixel areas in the original image are replaced by the superpixel. (1-P) percent remain unaltered.
ChangeColorspace(T, F, A)	Converts images from colorspace F to T and mixes with the original image using alpha A. Grayscale remains at three channels. (Fairly untested augmenter, use at own risk.)
Grayscale(A, F)	Converts images from colorspace F (default: RGB) to grayscale and mixes with the original image using alpha A.
GaussianBlur(S)	Blurs images using a gaussian kernel with size S.
AverageBlur(K)	Blurs images using a simple averaging kernel with size K.
MedianBlur(K)	Blurs images using a median over neihbourhoods of size K.
Convolve(M)	Convolves images with matrix M, which can be a lambda function.
Sharpen(A, L)	Runs a sharpening kernel over each image with lightness L (low values result in dark images). Mixes the result with the original image using alpha A.
Emboss(A, S)	Runs an emboss kernel over each image with strength S. Mixes the result with the original image using alpha A.
EdgeDetect(A)	Runs an edge detection kernel over each image. Mixes the result with the original image using alpha A.
DirectedEdgeDetect(A, D)	Runs a directed edge detection kernel over each image, which detects each from direction D (default: random direction from 0 to 360 degrees, chosen per image). Mixes the result with the original image using alpha A.
Add(V, PCH)	Adds value V to each image. If PCH is true, then the the sampled values may be different per channel.
AddElementwise(V, PCH)	Adds value V to each pixel. If PCH is true, then the the sampled values may be different per channel (and pixel).
AdditiveGaussianNoise(L, S, PCH)	Adds white/gaussian noise pixelwise to an image. The noise comes from the normal distribution N(L,S). If PCH is true, then the sampled values may be different per channel (and pixel).
Multiply(V, PCH)	Multiplies each image by value V, leading to darker/brighter images. If PCH is true, then the the sampled values may be different per channel.
MultiplyElementwise(V, PCH)	Multiplies each pixel by value V, leading to darker/brighter pixels. If PCH is true, then the the sampled values may be different per channel (and pixel).
Dropout(P, PCH)	Sets pixels to zero with probability P. If PCH is true, then channels may be treated differently, otherwise whole pixels are set to zero.
CoarseDropout(P, SPX, SPC, PCH)	Like Dropout, but samples the locations of pixels that are to be set to zero from a coarser/smaller image, which has pixel size SPX or relative size SPC. I.e. if SPC has a small value, the coarse map is small, resulting in large rectangles being dropped.
Invert(P, PCH)	Inverts with probability P all pixels in an image, i.e. sets them to (1-pixel_value). If PCH is true, each channel is treated individually (leading to only some channels being inverted).
ContrastNormalization(S, PCH)	Changes the contrast in images, by moving pixel values away or closer to 128. The direction and strength is defined by S. If PCH is set to true, the process happens channel-wise with possibly different S.
Affine(S, TPX, TPC, R, SH, O, M, CVAL)	Applies affine transformations to images. Scales them by S (>1=zoom in, <1=zoom out), translates them by TPX pixels or TPC percent, rotates them by R degrees and shears them by SH degrees. Interpolation happens with order O (0 or 1 are good and fast). Areas can appear in the resulting image, which have no corresponding area in the original image. M defines, how to handle these. If M='constant' then CVAL defines a constant value with which to fill the area.
PiecewiseAffine(S, R, C, O, M, CVAL)	Places a regular grid of points on the image. The grid has R rows and C columns. Then moves the points (and the image areas around them) by amounts that are samples from normal distribution N(0,S), leading to local distortions of varying strengths. O, M and CVAL are defines as in Affine.
ElasticTransformation(S, SM)	Moves each pixel individually around based on distortion fields. SM defines the smoothness of the distortion field and S its strength.