Multi-Constraint Adversarial Networks for Unsupervised Image-to-Image Translation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised image-to-image translation learns mappings without paired data.
• Generative Adversarial Networks (GANs) have advanced image translation.
• Existing GANs struggle with training instability due to large domain discrepancies.

Purpose of the Study:

• To propose a novel multi-constraint adversarial model (MCGAN) for robust unsupervised image translation.
• To address training instability and improve gradient flow for large appearance/shape differences.

Main Methods:

• Introduced a multi-constraint adversarial model (MCGAN).
• Applied multiple adversarial constraints to generator's multi-scale outputs via a single discriminator.
• Utilized dense dilated convolutions in the discriminator for enhanced information flow.

Main Results:

• MCGAN successfully captures large discrepancies in appearance between domains.
• Improved training stability and reduced blurriness in generated images.
• Achieved state-of-the-art performance on cat-to-dog, horse-to-zebra, and apple-to-orange datasets.

Conclusions:

• The proposed MCGAN effectively enhances unsupervised image-to-image translation.
• The multi-scale constraint and dilated convolutions significantly improve stability and quality.
• MCGAN offers a robust solution for challenging image translation tasks.