Augmentor

Augmentor is an image-augmentation library designed to render the process of artificial dataset enlargement more convenient, less error prone, and easier to reproduce. This is achieved using probabilistic transformation pipelines.

Package Status	Package Evaluator	Build Status

Augmentor.jl is the Julia package for Augmentor. You can find the Python version here.

Introduction

The following code snipped shows how such a pipeline can be specified using simple building blocks. To show the effect we compiled a few resulting output images into a gif while using an example image from the ISIC archive as input.

using Augmentor, ISICArchive

# Define a pipeline
pipeline = [Rotate90(.5), Rotate270(.5), FlipX(.5), FlipY(.5), RCropSize(160, 160), Resize(64, 64)]

# Load an input image
input_img = get(ISICArchive.ImageThumbnailRequest(id = "5592ac599fc3c13155a57a85"))

# Apply pipeline on image
output_img = transform(pipeline, input_img)

Input	Output

Documentation

Check out the latest documentation

Additionally, you can make use of Julia's native docsystem. The following example shows how to get additional information on DirImageSource within Julia's REPL:

?DirImageSource

Installation

To install Augmentor.jl, start up Julia and type the following code snipped into the REPL. It makes use of the native Julia package manager. Once installed the Augmentor package can be imported just as any other Julia package.

Pkg.clone("git@github.com:Evizero/Augmentor.jl.git")
using Augmentor

Additionally, for example if you encounter any sudden issues, you can manually choose to be on the latest (untagged) development version.

Pkg.checkout("Augmentor")

Overview

There are three basic building-blocks to create a system for image augmentation, namely an image source, image transformations, and an image operation pipeline.

• ImageSouce: Functionality to access images from some data source, such as a directory.

• Pipeline: A chain or tree of (probabilistic) transformations that should be applied to a given image, or set of images.

• ImageTransformation: As the name suggests concrete subclasses define a specific transformation that can be applied to an image, or a set of images. Operations can also be lifted into a Either, which have a random probability of occuring, depending on the hyperparameter chance.

Examples

The following code shows how to create a simple probabilistic pipeline, which can be sampled on the fly:

# load an example image
using TestImages
img = testimage("toucan")

# create empty pipeline
pl = Array{ImageTransformation,1}()

# add transformations to pipeline
push!(pl, FlipX(0.5))     # lifted to Either(FlipX(), NoOp()). 50% chance of occuring
push!(pl, FlipY())        # not lifted. will always occur
push!(pl, CropRatio(1.5)) # crop out biggest area that satisfies aspect ration
push!(pl, Resize(64,64))

# transform example image
img_new = transform(pl, img)

RGBA Images.Image with:
  data: 64x64 Array{ColorTypes.RGBA{FixedPointNumbers.UFixed{UInt8,8}},2}
  properties:
    transformations:  Flip x-axis. Flip y-axis. Crop to 1.5 aspect ratio. Resize to 64x64.
    imagedescription: <suppressed>
    spatialorder:  x y
    pixelspacing:  1 1

Input (img)	Output (img_new)

You can also use an ImageSource to sample input data from

# define directory as an image source
src = DirImageSource("path/to/images/")

# randomly sample a few images from the source
imgs = rand(src, 5) # Vector{Image}

# transform each image independently
imgs_new = transform(pl, imgs)

# create the image source by indexing the directory's content.
# Note: this command won't actually load any images.
src = DirImageSource("mydatadir/")

# number of available images
n = length(src)

# images can be indexed like an array
img1 = src[1]

# loop through all images in the source
for img in src
    println(img)
end

It is also possible to create a pipeline in a more concise way

pl = [FlipX(.5), FlipY(.5), Resize(32,32)]

3-element Array{Augmentor.ImageTransformation,1}:
 Either: (50%) Flip x-axis. (50%) No transformation.
 Either: (50%) Flip y-axis. (50%) No transformation.
 Resize to 32x32.

License

This code is free to use under the terms of the MIT license.

Acknowledgment

This package makes heavy use of the following packages in order to provide it's main functionality. To see at full list of utilized packages, please take a look at the REQUIRE file.

• Images.jl
• Interpolations.jl
• AffineTransforms.jl
• PiecewiseAffineTransforms.jl