Akademik Veri Yönetim Sistemi
7. Uluslararası Uygulamalı İstatistik Kongresi, İstanbul, Türkiye, 11 - 13 Mayıs 2026, ss.1, (Özet Bildiri)
This study explores deep learning-based image-to-image translation, which focuses on learning mappings between distinct visual domains and has become an important area in computer vision due to applications such as style transfer and domain adaptation. Unpaired image translation is particularly important because many real-world datasets lack directly corresponding image pairs, making supervised learning approaches inapplicable.
To address this issue, CycleGAN is used to learn bidirectional mappings between two domains through cycle-consistency loss, which ensures that translated images can be mapped back to their original form while preserving structural information across domains. The model is composed of two generators and two discriminators that improve the realism and domain consistency of generated images through adversarial training.
In this work, the horse-to-zebra dataset is used to evaluate the model’s ability to capture domain-specific visual characteristics, particularly texture-level transformations such as zebra stripe patterns, while maintaining global structural content. A CycleGAN model is implemented and trained to assess its performance under limited computational resources and to observe its learning behavior in a controlled experimental setting.
Experimental results show that the model produces visually coherent translations while preserving the structural integrity of the input images. These findings highlight the effectiveness of CycleGAN for unpaired image-to-image translation tasks and demonstrate its applicability in practical deep learning-based vision problems.