Object-stable unsupervised dual contrastive learning image-to-image translation with query-selected attention and...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Contrastive learning is increasingly popular for unsupervised image-to-image (I2I) translation.
• Previous query-selected attention (QS-Attn) modules focused on maximizing mutual information but often overemphasized background elements due to selecting queries with similar significance.
• This limitation hinders the precise translation of salient objects.

Purpose of the Study:

• To propose a novel dual-learning framework, the object-stable dual contrastive learning generative adversarial network (OS-DCLGAN), to address the limitations of existing unsupervised I2I translation methods.
• To enhance the focus on significant domain-specific information for improved translation quality.
• To refine intermediate features by effectively emphasizing or suppressing them.

Main Methods:

• Introduced a dual-learning framework integrating QS-Attn with a convolutional block attention module (CBAM).
• Integrated CBAM before the QS-Attn module to refine intermediate features and capture significant domain information.
• Utilized CBAM's ability to learn emphasis and suppression for feature refinement.

Main Results:

• The proposed OS-DCLGAN framework demonstrated superior performance compared to recent approaches in various I2I translation tasks.
• The integration of CBAM effectively refined features, leading to improved translation accuracy and object stability.
• The framework showed significant effectiveness and versatility across different I2I translation scenarios.

Conclusions:

• The OS-DCLGAN framework offers an effective solution for unsupervised image-to-image translation by improving feature representation.
• The proposed method successfully overcomes the background overemphasis issue seen in previous QS-Attn modules.
• The framework's enhanced performance highlights its potential for various computer vision applications requiring precise image translation.