WY-NET: A NEW APPROACH to IMAGE SYNTHESIS with GENERATIVE ADVERSARIAL NETWORKS
| dc.contributor.author | Şahin, Emrullah | |
| dc.contributor.author | Talu, Muhammed Fatih | |
| dc.date.accessioned | 2024-08-04T19:53:09Z | |
| dc.date.available | 2024-08-04T19:53:09Z | |
| dc.date.issued | 2022 | |
| dc.department | İnönü Üniversitesi | en_US |
| dc.description.abstract | Conditional image synthesis is the translation of images from different domains with the same dimensions into each other. Generative Adversarial Networks (GANs) are commonly used in translation studies in this field. With the classical GAN approach, data are transferred between the encoder and decoder of the generator network in the image translation. While this data transfer increases the quality of the translated image, it also leads to data dependency. This dependence has two negative effects: First, it prevents the understanding of whether the encoder or the decoder causes the error in the translated images, other causes the image synthesis quality to depend on the parameter increment of the network. In this study, two different architectures (dY-Net, uY-Net) are proposed. These architectures are developed on the principle of equalizing high-level feature parameters in cross-domain image translation. The first of these architectures concentrates on the speed of image synthesis, the other on its quality. There is a significant reduction in data dependency and parameter space in the dY-Net architecture, which concentrates on speed performance in image synthesis. The uY-Net architecture, which concentrates on image synthesis quality, attempts to maximize the results of metrics that measure quality like SSIM and PSNR. Three different datasets (Maps, Cityscapes, and Denim2Mustache) were used for performance testing of the proposed architectures and existing image synthesizing approaches. As a result of the tests, it has been seen that the proposed architecture synthesized images with similar accuracy, although it has approximately 66% parameters compared to DiscoGAN, which is one of the existing approaches. The results obtained show that WY-Net architectures, which provide high performance and translation quality, can be used in image synthesis. | en_US |
| dc.identifier.endpage | 290 | en_US |
| dc.identifier.issn | 2687-6167 | |
| dc.identifier.issue | 050 | en_US |
| dc.identifier.startpage | 270 | en_US |
| dc.identifier.trdizinid | 1127605 | en_US |
| dc.identifier.uri | https://search.trdizin.gov.tr/yayin/detay/1127605 | |
| dc.identifier.uri | https://hdl.handle.net/11616/89500 | |
| dc.indekslendigikaynak | TR-Dizin | en_US |
| dc.language.iso | en | en_US |
| dc.relation.ispartof | Journal of scientific reports-A (Online) | en_US |
| dc.relation.publicationcategory | Makale - Ulusal Hakemli Dergi - Kurum Öğretim Elemanı | en_US |
| dc.rights | info:eu-repo/semantics/openAccess | en_US |
| dc.title | WY-NET: A NEW APPROACH to IMAGE SYNTHESIS with GENERATIVE ADVERSARIAL NETWORKS | en_US |
| dc.type | Article | en_US |
