Exploring Generative Adversarial Networks (GANs) for Image Synthesis

Abstract

Background: Generative Adversarial Networks (GANs) have revolutionized image synthesis by enabling the generation of highly realistic images. However, the effectiveness of GANs varies depending on architecture and training methodologies. This study evaluates the performanceof aproposed GAN model compared to advanced architectures such as Style GAN2 and BigGAN, using the CIFAR-10 dataset as a benchmark. Objective:The primary objective of this study is to assess the ability of GANs to generate high-quality, diverse images by comparing the proposed GANmodel with established architectures. Performance is evaluated using key metrics such as the Inception Score (IS) and Fréchet Inception Distance(FID).Method:An experimental study design was employed, utilizing a GAN architecture comprising a generator and a discriminator trained in an adversarial manner. The CIFAR-10 dataset, consisting of 60,000 images across 10 categories, was used for training and evaluation. The Inception Score (IS) and Fréchet Inception Distance (FID) were calculated to assess image quality and diversity. Subjective visual assessments and computational efficiency were also analyzed.Results:The proposed GAN achieved an IS of 7.8 and an FID of 25.5, indicating moderate image quality and diversity. In comparison, Style GAN2 and BigGAN out performed the proposed model with IS scores of 8.3 and 8.7, and FID scores of 15.2and14.0, respectively. Despite its lower performance in image synthesis, the proposed GAN exhibited a significantly reduced training time (36hours)compared to Style GAN2 (72 hours) and BigGAN (96 hours). No significant mode collapse was observed across the models. However, subjective evaluations confirmed that the proposed GAN produced images of lower visual quality than its counterparts.Conclusion: While the proposed GAN demonstrated efficient training times, it lagged in terms of image qualityand diversity compared to more advanced models. Future research should focus on optimizing training strategies and architectural improvements to enhance GAN performance while maintaining computational efficiency. Keywords: Generative Adversarial Networks (GANs), Image Synthesis, DeepLearning, Inception Score (IS), Fréchet Inception Distance (FID)