Discolight

Flashy, ravey and state-of-the-art image augmentations to boost the performance of deep convolutional neural networks

At a Glance

• The library works with images in HxWxC (channels last) format.
• Based on numpy, OpenCV, paperspace, imgaug and albumentation.
• Diversified set of augmentations available.
• Supports python 3.6-3.8
• Supports transformations on images, and bounding boxes.
• The library works with images in HxWxC format.

Setup

The easiest way to get started is to install the discolight package via pip:

$ pip install discolight

Alternatively, if you have cloned this repository, you can run pip on the contents of the repository folder:

$ pip install .

Supported Augmentations

	Sample Image	Augmented Image	Sample Image (Bounding Boxes)	Augmented Image (Bounding Boxes)
ColorTemperature Changes the color temperature of the input image.
GaussianNoise Add gaussian noise to the given image.
GrayScale Return a grayscale version of the given image.
HorizontalFlip Horizontally flips the given image.
ImageCompression Apply a compression effect to the given image.
MotionBlur Add motionblur to a given image.
OneOf Perform a randomly selected augmentation on the given image.
RandomCrop Randomly crops the given image.
RandomEraser Randomly erase a rectangular area in the given image.
RandomHSV Randomly shift the color space of the given image.
RandomRotate Randomly rotate the given image.
RandomScale Randomly scale the given image.
RandomShear Randomly shear the given image.
RandomTranslate Randomly Translate the given image.
Resize Resize an image without preserving aspect ratio.
ResizeMaintainAspectRatio Resize an image while preserving aspect ratio.
Rotate Rotate the given image.
SaltAndPepperNoise Add salt and pepper or RGB noise to the given image.
Scale Scale the given image.
Sepia Returns a given image passed through the sepia filter.
Sequence Perform a sequence of augmentations on the given image.
Shear Horizontally shear the given image.
Translate Translate the given image.
VerticalFlip Vertically flip the given image.

Generating Augmented Images on the Command Line

To augment a set of images with annotations, you must first prepare your input data and prepare YAML configuration file to describe the augmentations you want to perform:

input:
images:
loader: Directoryoptions:
directory: ./sample_images # The base path of where the images are storedannotations:
loader: FourCornersCSVoptions:
annotations_file: ./sample_images/annotations.csv # The CSV file with annotationsnormalized: trueoutput:
images:
writer: Directoryoptions:
directory: ./output_folderannotations:
writer: FourCornersCSVoptions:
annotations_file: ./output_folder/aug_annotations.csvnormalized: trueaugmentations:
- name: GrayScaleoptions:
probs: 0.5
- name: Rotateoptions:
probs: 0.7angle: 35
- name: GaussianNoiseoptions:
probs: 0.9
- name: Shearoptions: {}save-original: true # Whether to save the original images to the output foldersave-bbox: true # Whether bounding boxes should be drawn on the augmented images

The image and annotation loaders specify how the images and annotations should be loaded. The FourCornersCSV (so-named because it expects the bounding box to be specified with four coordinates) loader expects an input CSV file in the following format:

image_name,x_min,y_min,x_max,y_max,labelwheat1.jpg,0,0.799805,0.037109,0.94043,1wheat1.jpg,0.064453,0.892578,0.181641,0.949219,1
...

The chosen image loader will be used to load the images specified under the image_name column (for example, wheat1.jpg). In this case, the Directory image loader will fetch the images from the specified folder. For more information on available image and annotation loaders, see .

Now you can run Discolight:

$ discolight generate configuration.yml

You can also pass your YAML configuration on standard input:

$ cat configuration.yml | discolight generate

After generation is complete, the augmented images and annotations will be written out according to your configuration options. Here, we choose to save the annotations to a CSV file using the same four-corners format, and the augmented images will appear in the output_folder/ directory.

Library Functionality

In addition to invoking Discolight from the command line, you can also invoke it as a library. First, construct the augmentation you want to apply to an image. For example, here is a simple augmentation sequence:

fromdiscolight.discoimportdiscofromdiscolight.annoationsimportBoundingBoxfromdiscolight.util.imageimportload_image, save_imageseq=disco.Sequence(augmentations=[
disco.GaussianNoise(),
disco.RandomEraser(),
disco.Rotate(angle=35)
])

Then, load the image and specify your annotations:

image=load_image("path/to/image.jpg")
annotations= [
# BoundingBox(x_min, y_min, x_max, y_max, class_idx)BoundingBox(0, 819, 38, 963, 1),
BoundingBox(66, 914, 186, 972, 1),
BoundingBox(52, 526, 126, 645, 1),
]

Now, you can invoke seq as a function to apply your augmentation:

aug_img, aug_annotations=seq(image, annotations)

If you have no annotations, you can also apply seq only to an image:

aug_img=seq(image)

To save the augmented image, use the provided save_image function. You can optionally pass the augmented annotations (also an array of BoundingBox objects) to be drawn in red over the augmented image:

save_image("aug-with-bboxes.jpg", aug_img, annotations=aug_annotations)
save_image("aug-no-bboxes.jpg", aug_img)

If you prefer to perform additional operations on the image after augmentation, augmented images are OpenCV images in RGB format.

Development

To learn more about how to develop Discolight (e.g., adding additional augmentatations), see the guide in doc/develop.md.