Nov.2022 Artificial Intelligence, Home assignment
- Max Grönlund
- Joona Hytönen
- Ville Juuti
- Juho Kemppainen
- Yusuke Mikami
- Simo Turunen
- Go file "AI-2022-genart/Joensuu-cycleGAN.ipynb"
- Run notebook "AI-2022-genart/Joensuu-cycleGAN.ipynb"
Aim of this group project was to implement General adversarial networks (GAN) model architecture in our work where we suppose to transform pictures of our city (Joensuu) taken by us into art style like. GAN concept based on combination of two separated neural network. We used technique named CycleGAN image-to-image translation. This technique provides several different ways to manipulate images. We used the one which translates real photographs into specific art style.
We used Deepnote for collaborator purposes. We pulled original CycleGAN code from Github repository to our Deepnote notebook, then for image transformation we first implemented pre-trained art-style models like Monet, Ukiyoe, Cezanne and Vangogh and then trained models with training data proposed by CycleGAN developing team. Image manipulation of real photographs was succeeded. The number of real images is 14. Collection of real Images and results can be found in the 6. Experiments and results section.