MetalGAN: Multi-domain label-less image synthesis using cGANs and meta-learning.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image synthesis is a key research area in computer vision and deep learning.
• Current methods often require separate models for different image domains or attributes.
• A single, flexible architecture for multi-domain image generation is highly desirable.

Purpose of the Study:

• To propose a novel architecture and training algorithm for multi-domain image synthesis using a single network.
• To enable label-less domain transfer without hard-coded class labels.
• To develop a more flexible and robust image generation approach.

Main Methods:

• Developed MetalGAN, a novel architecture combining conditional Generative Adversarial Networks (cGAN) with Meta-Learning.
• Employed a small dataset portion and avoided hard-coded labels.
• Utilized Meta-Learning for efficient domain switching within the network.

Main Results:

• Demonstrated the capability of a single network to produce multi-domain outputs.
• Successfully performed label-less multi-domain image synthesis.
• Validated the approach on facial attribute transfer tasks using the CelebA dataset.

Conclusions:

• MetalGAN effectively addresses the multi-domain, label-less image synthesis problem.
• The proposed approach offers greater flexibility and robustness compared to traditional methods.
• This work advances the field of generative models for diverse image generation tasks.