Image Translation by Domain-Adversarial Training.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Image translation is crucial for computer graphics and vision, mapping input images to synthetic counterparts.
• Conditional generative adversarial networks (cGANs) have advanced image translation but typically require expensive supervised data.
• The scarcity of paired data limits the application of existing cGAN-based methods.

Purpose of the Study:

• To develop a versatile framework for image translation applicable to both supervised and unsupervised settings.
• To enhance the quality and fidelity of generated synthetic images, particularly in data-scarce scenarios.
• To reduce the mapping space and improve generation quality by learning image priors.

Main Methods:

• A common framework for image translation is proposed, adaptable for unsupervised learning.
• Conditioning the discriminator on unaligned targets to learn image priors and reduce the mapping space.
• Domain-adversarial training, inspired by domain adaptation, is employed to capture discriminative features for improved fidelity.

Main Results:

• The proposed method demonstrates effectiveness in image translation tasks.
• Experimental results show compelling performance compared to existing baseline methods.
• The framework successfully improves generation quality and fidelity.

Conclusions:

• The developed framework offers a flexible approach to image translation, effective in both supervised and unsupervised scenarios.
• The method's generality is highlighted by its adaptability to diverse tasks.
• This work advances unsupervised image translation by improving generation quality and feature representation.