Multi-Domain Image-to-Image Translation via a Unified Circular Framework

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image-to-image translation models, often based on Generative Adversarial Networks (GANs), typically focus on one-to-one domain mappings.
• Existing methods often overlook the valuable correlations and shared information present across different domain pairs.

Purpose of the Study:

• To develop a unified framework for multiple image-to-image translation domains.
• To leverage shared knowledge across domains to improve translation quality and enable mutual learning.

Main Methods:

• Proposes a unified circular framework incorporating a shared knowledge module for multi-domain translation.
• Introduces an absolute consistency loss applied to feature maps to ensure intra-domain consistency.
• The model is designed for end-to-end training.

Main Results:

• Demonstrates effectiveness across diverse translation scenarios, including Thermal IR switching, weather changing, and semantic transfer.
• The shared knowledge module facilitates complementary information sharing between translation pairs.
• Ensures intra-domain consistency through the proposed absolute consistency loss.

Conclusions:

• The unified circular framework effectively addresses limitations of one-to-one translation by exploiting inter-domain correlations.
• The proposed method enhances mutual learning and achieves state-of-the-art results in complex image-to-image translation tasks.