Design of single image generation models based on generative adversarial networks

Image generation from a single natural image using generative adversarial networks (GANs) has attracted extensive attention recently due to the GANs' practical ability to produce photorealistic images and their potential applications in computer vision. In this thesis, two novel Unconditional GAN models are developed and implemented to generate diverse, high-quality, realistic image samples from a single natural training image. In both the generator and discriminator neural networks, Model 1 incorporates self-attention mechanisms, densely connected block architecture, and relativistic average least square GAN with gradient penalty loss component to proficiently learn the characteristics of a single training image. Model 2 effectively learns the latent representation information of the single training image. A novel training process has been developed for Model 2 by utilizing the learned latent representation. To test the performance and demonstrate the capabilities of the developed GAN models, quantitative and qualitative comparisons were carried out with well-known single-image GAN models in the literature. The comparisons were conducted using quantitative metrics measuring image quality and diversity on datasets such as Places, LSUN, and ImageNet, while qualitative user studies were conducted to assess image realism. In addition, the developed GAN models are tested in image manipulation applications. Experimental findings indicate that the developed GAN models exhibit efficient parameter utilization, mitigate overfitting, and successfully learn the internal patch statistics of training images characterized by complex structures and textures, comparable performance in single-image generation tasks. User studies confirm that images synthesized by the developed GAN models are frequently mistaken for the single training image. In this regard, the GAN models proposed within the scope of the thesis can be used as a powerful tool for various image processing tasks as well as data augmentation in areas dealing with limited training data. April 2024, 141 pages. Keywords: Single Image Generation, Generative Adversarial Networks, Deep Learning, SelfAttention, DenseNet, ALI, Image Manipulation