Investigation of generative capacity related to DCGANs across varied discriminator architectures and parameter counts: A comparative study

Abstract

Generating lifelike images through generative models poses a significant challenge, where Generative Adversarial Networks (GANs), particularly Deep Convolutional GANs (DCGANs), are commonly employed for image synthesis. This study focuses on altering the DCGAN discriminators structure and parameter count, investigating their effects on the characteristics of the resulting generated images. Assessment of these models is carried out using the Frchet Inception Distance (FID) score, a metric that gauges the quality of generated image samples. The research specifically involves substituting some convolutional layers with fully-connected layers, and the ensuing outcomes are thoroughly compared to discern the impact of these structural changes. Furthermore, dropout was used to study the number of the parameters influence. This study compared the FID score of the models when the probability is 0, 0.2, 0.4, 0.6 and 0.8. Experimental results showed that the DCGAN with the fully-connected layers generated ability was stronger than the original one. Besides, when the probability of the dropout is 0.6, the images generated was the most realistic. Finally, the paper explained the possible reasons for the difference and proposed a better generative model based on DCGAN.