Generative adversarial network based on frequency domain data enhancement: Dual-way discriminator structure Copes...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generative adversarial networks (GANs) excel at image generation but typically require extensive datasets.
• Training GANs with limited data presents challenges such as overfitting and hinders broader application.
• Existing data augmentation methods often focus on pixel-level changes, neglecting image structure and contours.

Purpose of the Study:

• To develop a stable GAN training method for limited data scenarios.
• To improve the performance and robustness of GANs when data is scarce.
• To overcome the limitations of traditional data augmentation techniques.

Main Methods:

• A novel dual-discriminator network architecture was designed to mitigate overfitting in limited data conditions.
• An adaptive dynamic data augmentation strategy utilizing Laplace convolution kernels in the frequency domain was proposed.
• This frequency-domain augmentation implicitly increases training data without altering pixel space.

Main Results:

• The proposed FD-GAN demonstrated superior image generation capabilities compared to existing methods.
• The model achieved impressive Frechet Inception Distance (FID) scores: 4.58 on AFHQ-Cat, 12.007 on AFHQ-Dog, and 10.382 on TankDataSet.
• The dual-discriminator and frequency-domain augmentation effectively improved GAN performance under data constraints.

Conclusions:

• FD-GAN offers a robust solution for training GANs with limited datasets.
• The combination of a dual-discriminator and frequency-domain augmentation significantly enhances image generation quality and training stability.
• This research paves the way for wider adoption of GANs in data-scarce environments.