Exploring Generative Adversarial Networks for Image Synthesis in Computer Vision

EasyChair Preprint 12357

Abstract

This paper delves into the exploration of GANs for image synthesis, focusing on their underlying principles, diverse architectures, training methodologies, and practical applications. The paper begins by elucidating the foundational concepts of GANs, highlighting the adversarial training process wherein a generator network learns to generate synthetic images that are indistinguishable from real images, while a discriminator network learns to differentiate between real and synthetic images. It discusses the evolution of GAN architectures, from the seminal DCGAN to more advanced variants such as Style GAN and BigGAN, each offering unique capabilities and improvements in image synthesis quality and diversity. Furthermore, the paper explores various training techniques and optimization strategies employed in training GANs, including minibatch discrimination, spectral normalization, and progressive growing, aimed at stabilizing training and improving convergence. It also discusses challenges inherent in GAN training, such as mode collapse, gradient vanishing, and instability, along with recent advancements and solutions to address these challenges.

Keyphrases: computer, synthesis, vision

@booklet{EasyChair:12357,
  author    = {Julia Anderson and Chris Liu},
  title     = {Exploring Generative Adversarial Networks for Image Synthesis in Computer Vision},
  howpublished = {EasyChair Preprint 12357},
  year      = {EasyChair, 2024}}