Image-to-Image Translation Using GANs
With the recent advancement in deep learning techniques, the idea of domain transfer using image translation interested a number of researchers. The use of new deep learning networks called Generative Adversarial Networks (GANs) have made the image translation problem much simpler and intuitive. In GANs the basic machinery that performs the task is same as Convolutional Neural Networks, it consists of a Generator (that generates the image) and Discriminator (that tries to discriminate either the image it received is real or produced by any Generator).
Getting motivation from a recent arxiv paper ‘Unpaired Image-to-Image translation using cycle-consistent Adversarial Networks’, I have implemented a basic Generative Adversarial Network (GAN) on pytorch that translates digit ‘6’ from MNIST dataset to ‘9’. For Generator, I used a subset of CNN proposed in ’Convolutional Sketch Inversion’ paper. The architecture of the generator network is as follows:
With the recent advancement in deep learning techniques, the idea of domain transfer using image translation interested a number of researchers. The use of new deep learning networks called Generative Adversarial Networks (GANs) have made the image translation problem much simpler and intuitive. In GANs the basic machinery that performs the task is same as Convolutional Neural Networks, it consists of a Generator (that generates the image) and Discriminator (that tries to discriminate either the image it received is real or produced by any Generator).
Getting motivation from a recent arxiv paper ‘Unpaired Image-to-Image translation using cycle-consistent Adversarial Networks’, I have implemented a basic Generative Adversarial Network (GAN) on pytorch that translates digit ‘6’ from MNIST dataset to ‘9’. For Generator, I used a subset of CNN proposed in ’Convolutional Sketch Inversion’ paper. The architecture of the generator network is as follows:
For Discriminator, I have used LeNet architecture with two neurons in the last layer. A total of 5000 images are used for training. Unlike [1], I have only used the cross entropy loss to train my network. Here are few results of my network.
Results
Input Output Input Output
Results
Input Output Input Output
[1] Jun-Yan Zhu, Taesung Park, Phillip Isola and Alexei A. Efros, "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks", Arxiv, 2017