Codebox Software

Image Augmentation for Machine Learning in Python

Published: 16 Oct 2016

This is a simple data augmentation tool for image files, intended for use with machine learning data sets. The tool scans a directory containing image files, and generates new images by performing a specified set of augmentation operations on each file that it finds. This process multiplies the number of training examples that can be used when developing a neural network, and should significantly improve the resulting network's performance, particularly when the number of training examples is relatively small.

Run the utility from the command-line as follows:

python main.py <image dir> <transform1> <transform2> ...

The <image dir> argument should be the path to a directory containing the image files to be augmented. The utility will search the directory recursively for files with any of the following extensions: jpg, jpeg, bmp, png.

The <transform> arguments determine what types of augmentation operations will be performed, using the codes listed in the table below:

CodeDescriptionExample Values
fliph Horizontal Flip fliph
flipv Vertical Flip flipv
noise Adds random noise to the image noise_0.01
noise_0.5
rot Rotates the image by the specified amount rot_90
rot_-45
trans Shifts the pixels of the image by the specified amounts in the x and y directions trans_20_10
trans_-10_0
zoom Zooms into the specified region of the image, performing stretching/shrinking as necessary zoom_0_0_20_20
zoom_-10_-20_10_10
blur Blurs the image by the specified amount blur_1.5

Each transform argument results in one additional output image being generated for each input image. An argument may consist of one or more augmentation operations. Multiple operations within a single argument must be separated by commas, and the order in which the operations are performed will match the order in which they are specified within the argument.

Examples

Produce 2 output images for each input image, one of which is flipped horizontally, and one of which is flipped vertically

python main.py ./my_images fliph flipv

Produce 1 output image for each input image, by first rotating the image by 90° and then flipping it horizontally

python main.py ./my_images rot_90,fliph

Operations

Horizontal Flip

Mirrors the image around a vertical line running through its center

python main.py ./my_images fliph

Original Image Flipped Image

Vertical Flip

Mirrors the image around a horizontal line running through its center

python main.py ./my_images flipv

Original Image Flipped Image

Noise

Adds random noise to the image. The amount of noise to be added is specified by a floating-point numeric value that is included in the transform argument, the numeric value must be greater than 0.

python main.py ./my_images noise_0.01 noise_0.02 noise_0.05

Original Image Noisy Image Noisy Image Noisy Image

Rotate

Rotates the image. The angle of rotation, in degrees, is specified by a integer value that is included in the transform argument

python main.py ./my_images rot_90 rot_180 rot_-90

Original Image Rotated Image Rotated Image Rotated Image

Translate

Performs a translation on the image. The size of the translation in the x and y directions are specified by integer values that are included in the transform argument as follows:

trans_<x distance>_<y distance>

Positive distance values correspond to rightward and downward movement, negative values indicate the opposite.

python main.py ./my_images trans_20_20 trans_0_100

Original Image Translated Image Translated Image

Zoom/Stretch

Zooms in to (or out from) a particular area of the image. The top-left and bottom-right coordinates of the target region are specified by integer values included in the transform argument as follows:

zoom_<top-left x>_<top-left y>_<bottom-right x>_<bottom-right y>

By specifying a target region with an aspect ratio that differs from that of the source image, stretching transformations can be performed.

python main.py ./my_images zoom_150_0_300_150 zoom_0_50_300_150 zoom_200_0_300_300

Original Image Zoomed Image Stretched Image Stretched Image

Blur

Blurs the image. The amount of blurring is specified by a floating-point value included in the transform argument.

python main.py ./my_images blur_1.0 blur_2.0 blur_4.0

Original Image Blurred Image Blurred Image Blurred Image