SUGAN: A Stable U-Net Based Generative Adversarial Network.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) are key for image generation but struggle with quality-vs-stability trade-offs.
• U-Net based GANs (U-Net GANs) offer high-quality image synthesis but are prone to mode collapse.

Purpose of the Study:

• To propose a Stable U-Net GAN (SUGAN) to enhance training stability and image quality in GANs.
• To mitigate mode collapse and improve the definition of generated images.

Main Methods:

• Introduced a gradient normalization module to the U-Net GAN discriminator to reduce gradient magnitudes and prevent overfitting.
• Incorporated a modified residual network in the generator to improve image detail capture and edge definition.

Main Results:

• SUGAN demonstrated significantly improved Inception Score (IS) and Fréchet Inception Distance (FID) metrics compared to state-of-the-art GANs.
• Achieved stable training, higher quality, and greater diversity in generated images.
• Effectively addressed mode collapse and produced sharper image details.

Conclusions:

• The proposed SUGAN effectively balances image quality and training stability for GANs.
• SUGAN represents a significant advancement in image generation tasks, offering superior performance and reliability.
• The open-source release of SUGAN facilitates further research and application in the field.