Generative adversarial networks for image-to-image translation using disentangled representation | |
| Author | Harsha, Somaraju Sri |
| Call Number | AIT RSPR no.CS-22-03 |
| Subject(s) | Machine learning--Technological innovations Neural networks (Computer science) |
| Note | A research study submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Computer Science |
| Publisher | Asian Institute of Technology |
| Abstract | Unpaired Image-to-image translation is an exciting topic in computer vision. When an im age of one domain is given, an equivalent image in the target domain should be generated. To compute this task, we propose a method for unsupervised image-to-image translation from one domain to another domain using an approach called disentanglement. We make an assumption that a model can have two underlying representations called the content space (common across domains) and the attribute space (specific to each domain), and an im age can be transformed from one domain to another by translating the underlying attribute space. Our Sub-GAN model architecture learns this underlying mapping to transform the attribute space of one domain to another. Then, we used the original image’s content space and translated attribute space to generate an image in the target domain. We evaluated our architecture’s KID and FID scores and found that the model has comparable results with state-of-the-art models. |
| Year | 2022 |
| Type | Research Study Project Report (RSPR) |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Computer Science (CS) |
| Chairperson(s) | Chaklam Silpasuwanchai; |
| Examination Committee(s) | Dailey, Mathew N.;Mongkol Ekpanyapong; |
| Degree | Research Studies Project Report (M. Eng.) - Asian Institute of Technology, 2022 |