Related Experiment Video
Updated: Jun 28, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
529
Multi-attention bidirectional contrastive learning method for unpaired image-to-image translation
Benchen Yang1, Xuzhao Liu1, Yize Li1
1Software College, Liaoning Technical University, Huludao, China.
Plos One
|April 16, 2024
Summary
This study introduces MabCUT, a novel multi-attention bidirectional contrastive learning method for unpaired image-to-image translation. MabCUT effectively addresses feature variations and interdependencies, enhancing image quality and preserving edge details.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Unpaired image-to-image translation (I2IT) aims to map images between domains.
- Existing contrastive learning methods struggle with feature variations and interdependencies, leading to instability and blurred edges.
Purpose of the Study:
- To propose a new method, MabCUT, for unpaired I2IT that overcomes limitations of previous approaches.
- To improve model stability and preserve crucial image features during translation.
Main Methods:
- Developed a multi-attention bidirectional contrastive learning framework (MabCUT).
- Utilized domain-specific embedding blocks with depthwise separable convolutions.
- Implemented a pixel-level multi-attention extractor for feature selection.
- Incorporated depthwise separable convolutions in the generator for enhanced feature representation.
Main Results:
- MabCUT demonstrated improved quality in unpaired I2IT across three datasets.
- The method successfully preserved essential source domain features.
- Addressed and avoided mode collapse-related image blurring.
Conclusions:
- MabCUT offers a robust solution for unpaired I2IT.
- The proposed method enhances translation quality and feature preservation.
- This work contributes to more stable and accurate cross-domain image transformation.
