Simple Yet Effective Way for Improving the Performance of GAN

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) often struggle with discriminator guidance due to reliance on nonrobust features.
• This limitation hinders the generator's ability to produce high-quality, realistic images.

Purpose of the Study:

• To propose a novel method to enhance GAN performance by improving discriminator feature extraction.
• To address the issue of discriminators focusing on irrelevant features, thereby improving generator guidance.

Main Methods:

• Introduction of a Cascading Rejection (CR) module for the discriminator.
• The CR module iteratively extracts diverse, non-overlapping features using vector rejection operations.
• The method is designed for easy integration into existing GAN architectures with minimal overhead.

Main Results:

• The CR module effectively prevents the discriminator from focusing on nonmeaningful features.
• This leads to more effective guidance for the generator, producing images more similar to real data.
• Quantitative evaluations on datasets like CIFAR-10, CelebA, and LSUN show significant improvements in Frechet Inception Distance (FID).

Conclusions:

• The proposed CR module offers a simple yet effective enhancement for GANs and conditional GANs.
• It improves the diversity and visual fidelity of generated images.
• The method demonstrates broad applicability across various datasets with marginal training overhead.