Progressively Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised image-to-image translation leverages generative adversarial networks (GANs) for learning domain mappings from unpaired data.
• Current state-of-the-art methods struggle with significant shape transitions and diverse target instances, often producing visual artifacts.
• Existing attention-based models fail to adequately address geometric transformations in image translation tasks.

Purpose of the Study:

• To propose a novel framework, Progressive Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization (PRO-U-GAT-IT), for unsupervised image-to-image translation.
• To address the limitations of existing methods in handling challenging translation tasks involving substantial shape changes and diverse instances.

Main Methods:

• Development of the PRO-U-GAT-IT framework incorporating attention mechanisms and adaptive layer-instance normalization.
• Focus on enabling holistic shape transformations beyond low-level feature translation.
• Utilizing unpaired image data for learning source-to-target domain mappings.

Main Results:

• PRO-U-GAT-IT demonstrates superior performance in unsupervised image-to-image translation compared to existing state-of-the-art models.
• The framework successfully handles images requiring extensive geometric and shape modifications.
• Experimental validation across multiple datasets confirms the proposed approach's effectiveness.

Conclusions:

• PRO-U-GAT-IT offers a robust solution for complex unsupervised image-to-image translation tasks, particularly those with significant geometric variations.
• The novel framework overcomes limitations of previous attention-based and GANs approaches in handling shape transitions.
• The proposed method advances the field by enabling more accurate and versatile image translation.