Image domain conversion using CycleGAN

This example shows how to convert images from one domain into another using CycleGAN

CycleGAN is a GAN model that is generally used for the following purposes.

• Style transfer (images and paintings)
• Season conversion
• Day / night conversion
• Object transformation

The difference from Pix2Pix, which also perform image-image conversion, is that CycleGAN uses unsupervised learning, so there is no need for a paired image dataset. In this example, even with unsupervised learning, you can see the model convert the images by understanding whether the fruit was a whole one or a cut one.

Requirements

• MATLAB
• Deep Learning Toolbox
• Image Processing Toolbox
• Parallel Computing Toolbox

MATLAB version should be R2019b and later

Usage

The repository provides the following files:

• CycleGANExample.mlx — Example showing how to train the CycleGAN model
• generator.m — Function to create a CycleGAN generator network
• discriminator.m — Function to create a CycleGAN discriminator network
• cycleGanImageDatastore.m — Datastore to prepare batches of images for training
• cycleGAN_1000.mat - Pretrained model that converts apples to oranges and vice-versa

To run, open CycleGANExample.mlx and run the script. You can train the model or use the pretrained model by setting the doTraining flag to false.

Reference

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (Jun-Yan Zhu.etc, 2017)

Copyright 2019-2020 The MathWorks, Inc. Download a free MATLAB trial for Deep Learning